Posts Tagged 'Performance'

July 16, 2013

Riak Performance Analysis: Bare Metal v. Virtual

In December, I posted a MongoDB performance analysis that showed the quantitative benefits of using bare metal servers for MongoDB workloads. It should come as no surprise that in the wake of SoftLayer's Riak launch, we've got some similar data to share about running Riak on bare metal.

To run this test, we started by creating five-node clusters with Riak 1.3.1 on SoftLayer bare metal servers and on a popular competitor's public cloud instances. For the SoftLayer environment, we created these clusters using the Riak Solution Designer, so the nodes were all provisioned, configured and clustered for us automatically when we ordered them. For the public cloud virtual instance Riak cluster, each node was provisioned indvidually using a Riak image template and manually configured into a cluster after all had come online. To optimize for Riak performance, I made a few tweaks at the OS level of our servers (running CentOS 64-bit):

Noatime
Nodiratime
barrier=0
data=writeback
ulimit -n 65536

The common Noatime and Nodiratime settings eliminate the need for writes during reads to help performance and disk wear. The barrier and writeback settings are a little less common and may not be what you'd normally set. Although those settings present a very slight risk for loss of data on disk failure, remember that the Riak solution is deployed in five-node rings with data redundantly available across multiple nodes in the ring. With that in mind and considering each node also being deployed with a RAID10 storage array, you can see that the minor risk for data loss on the failure of a single disk in the entire solution would have no impact on the entire data set (as there are plenty of redundant copies for that data available). Given the minor risk involved, the performance increases of those two settings justify their use.

With all of the nodes tweaked and configured into clusters, we set up Basho's test harness — Basho Bench — to remotely simulate load on the deployments. Basho Bench allows you to create a configurable test plan for a Riak cluster by configuring a number of workers to utilize a driver type to generate load. It comes packaged as an Erlang application with a config file example that you can alter to create the specifics for the concurrency, data set size, and duration of your tests. The results can be viewed as CSV data, and there is an optional graphics package that allows you to generate the graphs that I am posting in this blog. A simplified graphic of our test environment would look like this:

Riak Test Environment

The following Basho Bench config is what we used for our testing:

{mode, max}.
{duration, 120}.
{concurrent, 8}.
{driver, basho_bench_driver_riakc_pb}.
{key_generator,{int_to_bin,{uniform_int,1000000}}}.
{value_generator,{exponential_bin,4098,50000}}.
{riakc_pb_ips, [{10,60,68,9},{10,40,117,89},{10,80,64,4},{10,80,64,8},{10,60,68,7}]}.
{riakc_pb_replies, 2}.
{operations, [{get, 10},{put, 1}]}.

To spell it out a little simpler:

Tests Performed

Data Set: 400GB
10:1 Query-to-Update Operations
8 Concurrent Client Connections
Test Duration: 2 Hours

You may notice that in the test cases that use SoftLayer "Medium" Servers, the virtual provider nodes are running 26 virtual compute units against our dual proc hex-core servers (12 cores total). In testing with Riak, memory is important to the operations than CPU resources, so we provisioned the virtual instances to align with the 36GB of memory in each of the "Medium" SoftLayer servers. In the public cloud environment, the higher level of RAM was restricted to packages with higher CPU, so while the CPU counts differ, the RAM amounts are as close to even as we could make them.

One final "housekeeping" note before we dive into the results: The graphs below are pulled directly from the optional graphics package that displays Basho Bench results. You'll notice that the scale on the left-hand side of graphs differs dramatically between the two environments, so a cursory look at the results might not tell the whole story. Click any of the graphs below for a larger version. At the end of each test case, we'll share a few observations about the operations per second and latency results from each test. When we talk about latency in the "key observation" sections, we'll talk about the 99th percentile line — 99% of the results had latency below this line. More simply you could say, "This is the highest latency we saw on this platform in this test." The primary reason we're focusing on this line is because it's much easier to read on the graphs than the mean/median lines in the bottom graphs.

Riak Test 1: "Small" Bare Metal 5-Node Cluster vs Virtual 5-Node Cluster

Servers

SoftLayer Small Riak Server Node
Single 4-core Intel 1270 CPU
64-bit CentOS
8GB RAM
4 x 500GB SATAII – RAID10
1Gb Bonded Network
Virtual Provider Node
4 Virtual Compute Units
64-bit CentOS
7.5GB RAM
4 x 500GB Network Storage – RAID10
1Gb Network
 

Results

Riak Performance Analysis

Riak Performance Analysis

Key Observations

The SoftLayer environment showed much more consistency in operations per second with an average throughput around 450 Op/sec. The virtual environment throughput varied significantly between about 50 operations per second to more than 600 operations per second with the trend line fluctuating slightly between about 220 Op/sec and 350 Op/sec.

Comparing the latency of get and put requests, the 99th percentile of results in the SoftLayer environment stayed around 50ms for gets and under 200ms for puts while the same metric for the virtual environment hovered around 800ms in gets and 4000ms in puts. The scale of the graphs is drastically different, so if you aren't looking closely, you don't see how significantly the performance varies between the two.

Riak Test 2: "Medium" Bare Metal 5-Node Cluster vs Virtual 5-Node Cluster

Servers

SoftLayer Medium Riak Server Node
Dual 6-core Intel 5670 CPUs
64-bit CentOS
36GB RAM
4 x 300GB 15K SAS – RAID10
1Gb Network – Bonded
Virtual Provider Node
26 Virtual Compute Units
64-bit CentOS
30GB RAM
4 x 300GB Network Storage
1Gb Network
 

Results

Riak Performance Analysis

Riak Performance Analysis

Key Observations

Similar to the results of Test 1, the throughput numbers from the bare metal environment are more consistent (and are consistently higher) than the throughput results from the virtual instance environment. The SoftLayer environment performed between 1500 and 1750 operations per second on average while the virtual provider environment averaged around 1200 operations per second throughout the test.

The latency of get and put requests in Test 2 also paints a similar picture to Test 1. The 99th percentile of results in the SoftLayer environment stayed below 50ms and under 400ms for puts while the same metric for the virtual environment averaged about 250ms in gets and over 1000ms in puts. Latency in a big data application can be a killer, so the results from the virtual provider might be setting off alarm bells in your head.

Riak Test 3: "Medium" Bare Metal 5-Node Cluster vs Virtual 5-Node Cluster

Servers

SoftLayer Medium Riak Server Node
Dual 6-core Intel 5670 CPUs
64-bit CentOS
36GB RAM
4 x 128GB SSD – RAID10
1Gb Network – Bonded
Virtual Provider Node
26 Virtual Compute Units
64-bit CentOS
30GB RAM
4 x 300GB Network Storage
1Gb Network
 

Results

Riak Performance Analysis

Riak Performance Analysis

Key Observations

In Test 3, we're using the same specs in our virtual provider nodes, so the results for the virtual node environment are the same in Test 3 as they are in Test 2. In this Test, the SoftLayer environment substitutes SSD hard drives for the 15K SAS drives used in Test 2, and the throughput numbers show the impact of that improved I/O. The average throughput of the bare metal environment with SSDs is between 1750 and 2000 operations per second. Those numbers are slightly higher than the SoftLayer environment in Test 2, further distancing the bare metal results from the virtual provider results.

The latency of gets for the SoftLayer environment is very difficult to see in this graph because the latency was so low throughout the test. The 99th percentile of puts in the SoftLayer environment settled between 500ms and 625ms, which was a little higher than the bare metal results from Test 2 but still well below the latency from the virtual environment.

Summary

The results show that — similar to the majority of data-centric applications that we have tested — Riak has more consistent, better performing, and lower latency results when deployed onto bare metal instead of a cluster of public cloud instances. The stark differences in consistency of the results and the latency are noteworthy for developers looking to host their big data applications. We compared the 99th percentile of latency, but the mean/median results are worth checking out as well. Look at the mean and median results from the SoftLayer SSD Node environment: For gets, the mean latency was 2.5ms and the median was somewhere around 1ms. For puts, the mean was between 7.5ms and 11ms and the median was around 5ms. Those kinds of results are almost unbelievable (and that's why I've shared everything involved in completing this test so that you can try it yourself and see that there's no funny business going on).

It's commonly understood that local single-tenant resources that bare metal will always perform better than network storage resources, but by putting some concrete numbers on paper, the difference in performance is pretty amazing. Virtualizing on multi-tenant solutions with network attached storage often introduces latency issues, and performance will vary significantly depending on host load. These results may seem obvious, but sometimes the promise of quick and easy deployments on public cloud environments can lure even the sanest and most rational developer. Some applications are suited for public cloud, but big data isn't one of them. But when you have data-centric apps that require extreme I/O traffic to your storage medium, nothing can beat local high performance resources.

-Harold

April 30, 2013

Big Data at SoftLayer: Riak

Big data is only getting bigger. Late last year, SoftLayer teamed up with 10Gen to launch a high-performance MongoDB solution, and since then, many of our customers have been clamoring for us to support other big data platforms in the same way. By automating the provisioning process of a complex big data environment on bare metal infrastructure, we made life a lot easier for developers who demanded performance and on-demand scalability for their big data applications, and it's clear that our simple formula produced amazing results. As Marc mentioned when he started breaking down big data database models, document-oriented databases like MongoDB are phenomenal for certain use-cases, and in other situations, a key-value store might be a better fit. With that in mind, we called up our friends at Basho and started building a high-performance architecture specifically for Riak ... And I'm excited to announce that we're launching it today!

Riak is an open source, distributed database platform based on the principles enumerated in the DynamoDB paper. It uses a simple key/value model for object storage, and it was architected for high availability, fault tolerance, operational simplicity and scalability. A Riak cluster is composed of multiple nodes that are all connected, all communicating and sharing data automatically. If one node were to fail, the other nodes would automatically share the data that the failed node was storing and processing until the node is back up and running or a new node is added. See the diagram below for a simple illustration of how adding a node to a cluster works within Riak.

Riak Nodes

We will support both the open source and the Enterprise versions of Riak. The open source version is a great place to start. It has all of the database functionality of Riak Enterprise, but it is limited to a single cluster. The Enterprise version supports replication between clusters across data centers, giving you lots of architectural options. You can use replication to build highly available, live-live failover applications. You can also use it to distribute your application's data across regions, giving you a global platform that you can update anywhere in the world and know that those modifications will be available anywhere else. Riak Enterprise customers also receive 24×7 coverage, both from SoftLayer and Basho. This includes SoftLayer's one-hour guaranteed response for Severity 1 hardware issues and unlimited support available via our secure web portal, email and phone.

The business use-case for this flexibility is that if you need to scale up or down, nodes can be easily added or taken down as your requirements change. You can opt for a single-data center environment with a few nodes or you can broaden your architecture to a multi-data center deployment with a 40-node cluster. While these capabilities are inherent in Riak, they can be complicated to build and configure, so we spent countless hours working with Basho to streamline Riak deployment on the SoftLayer platform. The fruit of that labor can be found in our Riak Solution Designer:

Riak Solution Designer

The server configurations and packages in the Riak Solution Designer have been selected to deliver the performance, availability and stability that our customers expect from their bare metal and virtual cloud infrastructure at SoftLayer. With a few quick clicks, you can order a fully configured Riak environment, and it'll be provisioned and online for you in two to four hours. And everything you order is on a month-to-month contract.

Thanks to the hard work done by the SoftLayer development group and Basho's team, we're proud to be the first in the marketplace to offer a turn-key Riak solution on bare metal infrastructure. You don't need to sacrifice performance and agility for simplicity.

For more information, visit SoftLayer.com/Riak or contact our sales team.

-Duke

January 24, 2013

Startup Series: SPEEDILICIOUS

Research from the Aberdeen Group shows the average website is losing 9% of its business because
 the speed of the site frustrates visitors into leaving. 9% of your traffic might be leaving your site because they feel like it's too slow. That thought is staggering, and any site owner would be foolish not to fix the problem. SPEEDILICIOUS — one of our new Catalyst partners — has an innovative solution that optimizes website performance and helps businesses deliver content to their end users faster.

SPEEDILICIOUS

I recently had the chance to chat with SPEEDILICIOUS founders Seymour Segnit and Chip Krauskopf, and Seymour rephrased that "9%" statistic in a pretty alarming way: "Losing 9% of your business is the equivalent of simply allowing your website to go offline, down, dark, dead, 404 for over a MONTH each year!" There is ample data to back this up from high-profile sites like Amazon, Microsoft and Walmart.com, but intuitively, you know it already ... A slow site (even a slightly slow site) is annoying.

The challenge many website owners have when it comes to their loading speeds is that problems might not be noticeable from their own workstations. Thanks to caching and the Internet connections most of us have, when we visit our own sites, we don't have any trouble accessing our content quickly. Unfortunately, many of our customers don't share that experience when they visit our sites on mobile, hotel, airports and (worst of all) conference connections. The most common approach to speeding up load times is to throw bigger servers or a CDN (content delivery network) at the problem, but while those improvements make a difference, they only address part of the problem ... Even with the most powerful servers in SoftLayer's fleet, your page can load at a crawl if your code can't be rendered quickly by a browser.

That makes life as a website developer difficult. The process of optimizing code and tweaking settings to speed up load times can be time-consuming and frustrating. Or as Chip explained to me, "Speeding up your site is essential, it shouldn’t be be slow and complicated. We fix that problem." Take a look:

The idea that your site performance can be sped up significantly overnight seems a little crazy, but if it works (which it clearly does), wouldn't it be crazier not to try it? SPEEDILICIOUS offers a $1 trial for you to see the results on your own site, and they regularly host a free webinar called "How to Grow Your Business 5-15% Overnight" which covers the critical techniques for speeding up any website.

As technology continues to improve and behavioral patterns of purchasing migrate away from the mall and onto our computers and smart phones, SPEEDILICIOUS has a tremendous opportunity to capture a ripe market. So they're clearly a great fit for Catalyst. If you're interested in learning more or would like to speak to Seymour, Chip or anyone on their team, please let me know and I'll make the direct introduction any time.

-@JoshuaKrammes

December 20, 2012

MongoDB Performance Analysis: Bare Metal v. Virtual

Developers can be cynical. When "the next great thing in technology" is announced, I usually wait to see how it performs before I get too excited about it ... Show me how that "next great thing" compares apples-to-apples with the competition, and you'll get my attention. With the launch of MongoDB at SoftLayer, I'd guess a lot of developers outside of SoftLayer and 10gen have the same "wait and see" attitude about the new platform, so I put our new MongoDB engineered servers to the test.

When I shared MongoDB architectural best practices, I referenced a few of the significant optimizations our team worked with 10gen to incorporate into our engineered servers (cheat sheet). To illustrate the impact of these changes in MongoDB performance, we ran 10gen's recommended benchmarking harness (freely available for download and testing of your own environment) on our three tiers of engineered servers alongside equivalent shared virtual environments commonly deployed by the MongoDB community. We've made a pretty big deal about the performance impact of running MongoDB on optimized bare metal infrastructure, so it's time to put our money where our mouth is.

The Testing Environment

For each of the available SoftLayer MongoDB engineered servers, data sets of 512kb documents were preloaded onto single MongoDB instances. The data sets were created with varying size compared to available memory to allow for data sets that were both larger (2X) and smaller than available memory. Each test also ensured that the data set was altered during the test run frequently enough to prevent the queries from caching all of the data into memory.

Once the data sets were created, JMeter server instances with 4 cores and 16GB of RAM were used to drive 'benchrun' from the 10gen benchmarking harness. This diagram illustrates how we set up the testing environment (click for a better look):

MongoDB Performance Analysis Setup

These Jmeter servers function as the clients generating traffic on the MongoDB instances. Each client generated random query and update requests with a ratio of six queries per update (The update requests in the test were to ensure that data was not allowed to fully cache into memory and never exercise reads from disk). These tests were designed to create an extreme load on the servers from an exponentially increasing number of clients until the system resources became saturated, and we recorded the resulting performance of the MongoDB application.

At the Medium (MD) and Large (LG) engineered server tiers, performance metrics were run separately for servers using 15K SAS hard drive data mounts and servers using SSD hard drive data mounts. If you missed the post comparing the IOPS statistics between different engineered server hard drive configurations, be sure to check it out. For a better view of the results in a given graph, click the image included in the results below to see a larger version.

Test Case 1: Small MongoDB Engineered Servers vs Shared Virtual Instance

Servers

Small (SM) MongoDB Engineered Server
Single 4-core Intel 1270 CPU
64-bit CentOS
8GB RAM
2 x 500GB SATAII - RAID1
1Gb Network
Virtual Provider Instance
4 Virtual Compute Units
64-bit CentOS
7.5GB RAM
2 x 500GB Network Storage - RAID1
1Gb Network
 

Tests Performed

Small Data Set (8GB of .5mb documents)
200 iterations of 6:1 query-to-update operations
Concurrent client connections exponentially increased from 1 to 32
Test duration spanned 48 hours
Average Read Operations per Second
by Concurrent Client
MongoDB Performance Analysis
Peak Read Operations per Second
by Concurrent ClientMongoDB Performance Analysis
Average Write Operations per Second
by Concurrent Client
MongoDB Performance Analysis
Peak Write Operations per Second
by Concurrent ClientMongoDB Performance Analysis

Test Case 2: Medium MongoDB Engineered Servers vs Shared Virtual Instance

Servers (15K SAS Data Mount Comparison)

Medium (MD) MongoDB Engineered Server
Dual 6-core Intel 5670 CPUs
64-bit CentOS
36GB RAM
2 x 64GB SSD - RAID1 (Journal Mount)
4 x 300GB 15K SAS - RAID10 (Data Mount)
1Gb Network - Bonded
Virtual Provider Instance
26 Virtual Compute Units
64-bit CentOS
30GB RAM
2 x 64GB Network Storage - RAID1 (Journal Mount)
4 x 300GB Network Storage - RAID10 (Data Mount)
1Gb Network
 

Tests Performed

Small Data Set (32GB of .5mb documents)
200 iterations of 6:1 query-to-update operations
Concurrent client connections exponentially increased from 1 to 128
Test duration spanned 48 hours
Average Read Operations per Second
by Concurrent Client
MongoDB Performance Analysis
Peak Read Operations per Second
by Concurrent ClientMongoDB Performance Analysis
Average Write Operations per Second
by Concurrent Client
MongoDB Performance Analysis
Peak Write Operations per Second
by Concurrent ClientMongoDB Performance Analysis

Servers (SSD Data Mount Comparison)

Medium (MD) MongoDB Engineered Server
Dual 6-core Intel 5670 CPUs
64-bit CentOS
36GB RAM
2 x 64GB SSD - RAID1 (Journal Mount)
4 x 400GB SSD - RAID10 (Data Mount)
1Gb Network - Bonded
Virtual Provider Instance
26 Virtual Compute Units
64-bit CentOS
30GB RAM
2 x 64GB Network Storage - RAID1 (Journal Mount)
4 x 300GB Network Storage - RAID10 (Data Mount)
1Gb Network
 

Tests Performed

Small Data Set (32GB of .5mb documents)
200 iterations of 6:1 query-to-update operations
Concurrent client connections exponentially increased from 1 to 128
Test duration spanned 48 hours
Average Read Operations per Second
by Concurrent Client
MongoDB Performance Analysis
Peak Read Operations per Second
by Concurrent ClientMongoDB Performance Analysis
Average Write Operations per Second
by Concurrent Client
MongoDB Performance Analysis
Peak Write Operations per Second
by Concurrent ClientMongoDB Performance Analysis

Test Case 3: Large MongoDB Engineered Servers vs Shared Virtual Instance

Servers (15K SAS Data Mount Comparison)

Large (LG) MongoDB Engineered Server
Dual 8-core Intel E5-2620 CPUs
64-bit CentOS
128GB RAM
2 x 64GB SSD - RAID1 (Journal Mount)
6 x 600GB 15K SAS - RAID10 (Data Mount)
1Gb Network - Bonded
Virtual Provider Instance
26 Virtual Compute Units
64-bit CentOS
64GB RAM (Maximum available on this provider)
2 x 64GB Network Storage - RAID1 (Journal Mount)
6 x 600GB Network Storage - RAID10 (Data Mount)
1Gb Network
 

Tests Performed

Small Data Set (64GB of .5mb documents)
200 iterations of 6:1 query-to-update operations
Concurrent client connections exponentially increased from 1 to 128
Test duration spanned 48 hours
Average Read Operations per Second
by Concurrent Client
MongoDB Performance Analysis
Peak Read Operations per Second
by Concurrent ClientMongoDB Performance Analysis
Average Write Operations per Second
by Concurrent Client
MongoDB Performance Analysis
Peak Write Operations per Second
by Concurrent ClientMongoDB Performance Analysis

Servers (SSD Data Mount Comparison)

Large (LG) MongoDB Engineered Server
Dual 8-core Intel E5-2620 CPUs
64-bit CentOS
128GB RAM
2 x 64GB SSD - RAID1 (Journal Mount)
6 x 400GB SSD - RAID10 (Data Mount)
1Gb Network - Bonded
Virtual Provider Instance
26 Virtual Compute Units
64-bit CentOS
64GB RAM (Maximum available on this provider)
2 x 64GB Network Storage - RAID1 (Journal Mount)
6 x 600GB Network Storage - RAID10 (Data Mount)
1Gb Network
 

Tests Performed

Small Data Set (64GB of .5mb documents)
200 iterations of 6:1 query-to-update operations
Concurrent client connections exponentially increased from 1 to 128
Test duration spanned over 48 hours
Average Read Operations per Second
by Concurrent Client
MongoDB Performance Analysis
Peak Read Operations per Second
by Concurrent ClientMongoDB Performance Analysis
Average Write Operations per Second
by Concurrent Client
MongoDB Performance Analysis
Peak Write Operations per Second
by Concurrent ClientMongoDB Performance Analysis

Impressions from Performance Testing

The results speak for themselves. Running a Mongo DB big data solution on a shared virtual environment has significant drawbacks when compared to running MongoDB on a single-tenant bare metal offering. Disk I/O is by far the most limiting resource for MongoDB, and relying on shared network-attached storage (with much lower disk I/O) makes this limitation very apparent. Beyond the average and peak statistics above, performance varied much more significantly in the virtual instance environment, so it's not as consistent and predictable as a bare metal.

Highlights:

  • When a working data set is smaller than available memory, query performance increases.
  • The number of clients performing queries has an impact on query performance because more data is being actively cached at a rapid rate.
  • The addition of a separate Journal Mount volume significantly improves performance. Because the Small (SM) engineered server does not include a secondary mount for Journals, whenever MongoDB began to journal, the disk I/O associated with journalling was disruptive to the query and update operations performed on the Data Mount.
  • The best deployments in terms of operations per second, stability and control were the configurations with a RAID10 SSD Data Mount and a RAID1 SSD Journal Mount. These configurations are available in both our Medium and Large offerings, and I'd highly recommend them.

-Harold

December 17, 2012

Big Data at SoftLayer: The Importance of IOPS

The jet flow gates in the Hoover Dam can release up to 73,000 cubic feet — the equivalent of 546,040 gallons — of water per second at 120 miles per hour. Imagine replacing those jet flow gates with a single garden hose that pushes 25 gallons per minute (or 0.42 gallons per second). Things would get ugly pretty quickly. In the same way, a massive "big data" infrastructure can be crippled by insufficient IOPS.

IOPS — Input/Output Operations Per Second — measure computer storage in terms of the number of read and write operations it can perform in a second. IOPS are a primary concern for database environments where content is being written and queried constantly, and when we take those database environments to the extreme (big data), the importance of IOPS can't be overstated: If you aren't able perform database reads and writes quickly in a big data environment, it doesn't matter how many gigabytes, terabytes or petabytes you have in your database ... You won't be able to efficiently access, add to or modify your data set.

As we worked with 10gen to create, test and tweak SoftLayer's MongoDB engineered servers, our primary focus centered on performance. Since the performance of massively scalable databases is dictated by the read and write operations to that database's data set, we invested significant resources into maximizing the IOPS for each engineered server ... And that involved a lot more than just swapping hard drives out of servers until we found a configuration that worked best. Yes, "Disk I/O" — the amount of input/output operations a given disk can perform — plays a significant role in big data IOPS, but many other factors limit big data performance. How is performance impacted by network-attached storage? At what point will a given CPU become a bottleneck? How much RAM should included in a base configuration to accommodate the load we expect our users to put on each tier of server? Are there operating system changes that can optimize the performance of a platform like MongoDB?

The resulting engineered servers are a testament to the blood, sweat and tears that were shed in the name of creating a reliable, high-performance big data environment. And I can prove it.

Most shared virtual instances — the scalable infrastructure many users employ for big data — use network-attached storage for their platform's storage. When data has to be queried over a network connection (rather than from a local disk), you introduce latency and more "moving parts" that have to work together. Disk I/O might be amazing on the enterprise SAN where your data lives, but because that data is not stored on-server with your processor or memory resources, performance can sporadically go from "Amazing" to "I Hate My Life" depending on network traffic. When I've tested the IOPS for network-attached storage from a large competitor's virtual instances, I saw an average of around 400 IOPS per mount. It's difficult to say whether that's "not good enough" because every application will have different needs in terms of concurrent reads and writes, but it certainly could be better. We performed some internal testing of the IOPS for the hard drive configurations in our Medium and Large MongoDB engineered servers to give you an apples-to-apples comparison.

Before we get into the tests, here are the specs for the servers we're using:

Medium (MD) MongoDB Engineered Server
Dual 6-core Intel 5670 CPUs
CentOS 6 64-bit
36GB RAM
1Gb Network - Bonded
Large (LG) MongoDB Engineered Server
Dual 8-core Intel E5-2620 CPUs
CentOS 6 64-bit
128GB RAM
1Gb Network - Bonded
 

The numbers shown in the table below reflect the average number of IOPS we recorded with a 100% random read/write workload on each of these engineered servers. To measure these IOPS, we used a tool called fio with an 8k block size and iodepth at 128. Remembering that the virtual instance using network-attached storage was able to get 400 IOPS per mount, let's look at how our "base" configurations perform:

Medium - 2 x 64GB SSD RAID1 (Journal) - 4 x 300GB 15k SAS RAID10 (Data)
Random Read IOPS - /var/lib/mongo/logs 2937
Random Write IOPS - /var/lib/mongo/logs 1306
Random Read IOPS - /var/lib/mongo/data 1720
Random Write IOPS - /var/lib/mongo/data 772
Random Read IOPS - /var/lib/mongo/data/journal 19659
Random Write IOPS - /var/lib/mongo/data/journal 8869
   
Medium - 2 x 64GB SSD RAID1 (Journal) - 4 x 400GB SSD RAID10 (Data)
Random Read IOPS - /var/lib/mongo/logs 30269
Random Write IOPS - /var/lib/mongo/logs 13124
Random Read IOPS - /var/lib/mongo/data 33757
Random Write IOPS - /var/lib/mongo/data 14168
Random Read IOPS - /var/lib/mongo/data/journal 19644
Random Write IOPS - /var/lib/mongo/data/journal 8882
   
Large - 2 x 64GB SSD RAID1 (Journal) - 6 x 600GB 15k SAS RAID10 (Data)
Random Read IOPS - /var/lib/mongo/logs 4820
Random Write IOPS - /var/lib/mongo/logs 2080
Random Read IOPS - /var/lib/mongo/data 2461
Random Write IOPS - /var/lib/mongo/data 1099
Random Read IOPS - /var/lib/mongo/data/journal 19639
Random Write IOPS - /var/lib/mongo/data/journal 8772
 
Large - 2 x 64GB SSD RAID1 (Journal) - 6 x 400GB SSD RAID10 (Data)
Random Read IOPS - /var/lib/mongo/logs 32403
Random Write IOPS - /var/lib/mongo/logs 13928
Random Read IOPS - /var/lib/mongo/data 34536
Random Write IOPS - /var/lib/mongo/data 15412
Random Read IOPS - /var/lib/mongo/data/journal 19578
Random Write IOPS - /var/lib/mongo/data/journal 8835

Clearly, the 400 IOPS per mount results you'd see in SAN-based storage can't hold a candle to the performance of a physical disk, regardless of whether it's SAS or SSD. As you'd expect, the "Journal" reads and writes have roughly the same IOPS between all of the configurations because all four configurations use 2 x 64GB SSD drives in RAID1. In both configurations, SSD drives provide better Data mount read/write performance than the 15K SAS drives, and the results suggest that having more physical drives in a Data mount will provide higher average IOPS. To put that observation to the test, I maxed out the number of hard drives in both configurations (10 in the 2U MD server and 34 in the 4U LG server) and recorded the results:

Medium - 2 x 64GB SSD RAID1 (Journal) - 10 x 300GB 15k SAS RAID10 (Data)
Random Read IOPS - /var/lib/mongo/logs 7175
Random Write IOPS - /var/lib/mongo/logs 3481
Random Read IOPS - /var/lib/mongo/data 6468
Random Write IOPS - /var/lib/mongo/data 1763
Random Read IOPS - /var/lib/mongo/data/journal 18383
Random Write IOPS - /var/lib/mongo/data/journal 8765
   
Medium - 2 x 64GB SSD RAID1 (Journal) - 10 x 400GB SSD RAID10 (Data)
Random Read IOPS - /var/lib/mongo/logs 32160
Random Write IOPS - /var/lib/mongo/logs 12181
Random Read IOPS - /var/lib/mongo/data 34642
Random Write IOPS - /var/lib/mongo/data 14545
Random Read IOPS - /var/lib/mongo/data/journal 19699
Random Write IOPS - /var/lib/mongo/data/journal 8764
   
Large - 2 x 64GB SSD RAID1 (Journal) - 34 x 600GB 15k SAS RAID10 (Data)
Random Read IOPS - /var/lib/mongo/logs 17566
Random Write IOPS - /var/lib/mongo/logs 11918
Random Read IOPS - /var/lib/mongo/data 9978
Random Write IOPS - /var/lib/mongo/data 6526
Random Read IOPS - /var/lib/mongo/data/journal 18522
Random Write IOPS - /var/lib/mongo/data/journal 8722
 
Large - 2 x 64GB SSD RAID1 (Journal) - 34 x 400GB SSD RAID10 (Data)
Random Read IOPS - /var/lib/mongo/logs 34220
Random Write IOPS - /var/lib/mongo/logs 15388
Random Read IOPS - /var/lib/mongo/data 35998
Random Write IOPS - /var/lib/mongo/data 17120
Random Read IOPS - /var/lib/mongo/data/journal 17998
Random Write IOPS - /var/lib/mongo/data/journal 8822

It should come as no surprise that by adding more drives into the configuration, we get better IOPS, but you might be wondering why the results aren't "betterer" when it comes to the IOPS in the SSD drive configurations. While the IOPS numbers improve going from four to ten drives in the medium engineered server and six to thirty-four drives in the large engineered server, they don't increase as significantly as the IOPS differences in the SAS drives. This is what I meant when I explained that several factors contribute to and potentially limit IOPS performance. In this case, the limiting factor throttling the (ridiculously high) IOPS is the RAID card we are using in the servers. We've been working with our RAID card vendor to test a new card that will open a little more headroom for SSD IOPS, but that replacement card doesn't provide the consistency and reliability we need for these servers (which is just as important as speed).

There are probably a dozen other observations I could point out about how each result compares with the others (and why), but I'll stop here and open the floor for you. Do you notice anything interesting in the results? Does anything surprise you? What kind of IOPS performance have you seen from your server/cloud instance when running a tool like fio?

-Kelly

December 6, 2012

MongoDB: Architectural Best Practices

With the launch of our MongoDB solutions, developers can provision powerful, optimized, horizontally scaling NoSQL database clusters in real-time on bare metal infrastructure in SoftLayer data centers around the world. We worked tirelessly with our friends at 10gen — the creators of MongoDB — to build and tweak hardware and software configurations that enable peak MongoDB performance, and the resulting platform is pretty amazing. As Duke mentioned in his blog post, those efforts followed 10Gen's MongoDB best practices, but what he didn't mention was that we created some architectural best practices of our own for MongoDB in deployments on our platform.

The MongoDB engineered servers that you order from SoftLayer already implement several of the recommendations you'll see below, and I'll note which have been incorporated as we go through them. Given the scope of the topic, it's probably easiest to break down this guide into a few sections to make it a little more digestible. Let's take a look at the architectural best practices of running MongoDB through the phases of the roll-out process: Selecting a deployment strategy to prepare for your MongoDB installation, the installation itself, and the operational considerations of running it in production.

Deployment Strategy

When planning your MongoDB deployment, you should follow Sun Tzu's (modified) advice: "If you know the [friend] and know yourself, you need not fear the result of a hundred battles." "Friend" was substituted for the "enemy" in this advice because the other party is MongoDB. If you aren't familiar with MongoDB, the top of your to-do list should be to read MongoDB's official documentation. That information will give you the background you'll need as you build and use your database. When you feel comfortable with what MongoDB is all about, it's time to "know yourself."

Your most important consideration will be the current and anticipated sizes of your data set. Understanding the volume of data you'll need to accommodate will be the primary driver for your choice of individual physical nodes as well as your sharding plans. Once you've established an expected size of your data set, you need to consider the importance of your data and how tolerant you are of the possibility of lost or lagging data (especially in replicated scenarios). With this information in hand, you can plan and start testing your deployment strategy.

It sounds a little strange to hear that you should test a deployment strategy, but when it comes to big data, you want to make sure your databases start with a strong foundation. You should perform load testing scenarios on a potential deployment strategy to confirm that a given architecture will meet your needs, and there are a few specific areas that you should consider:

Memory Sizing
MongoDB (like many data-oriented applications) works best when the data set can reside in memory. Nothing performs better than a MongoDB instance that does not require disk I/O. Whenever possible, select a platform that has more available RAM than your working data set size. If your data set exceeds the available RAM for a single node, then consider using sharding to increase the amount of available RAM in a cluster to accommodate the larger data set. This will maximize the overall performance of your deployment. If you notice page faults when you put your database under production load, they may indicate that you are exceeding the available RAM in your deployment.

Disk Type
If speed is not your primary concern or if you have a data set that is far larger than any available in memory strategy can support, selecting the proper disk type for your deployment is important. IOPS will be key in selecting your disk type and obviously the higher the IOPS the better the performance of MongoDB. Local disks should be used whenever possible (as network storage can cause high latency and poor performance for your deployment). It's also advised that you use RAID 10 when creating disk arrays.

To give you an idea of what kind of IOPS to expect from a given type of drive, these are the approximate ranges of IOPS per drive in SoftLayer MongoDB engineered servers:

SATA II – 100-200 IOPS
15K SAS – 300-400 IOPS
SSD – 7,000-8,000 IOPS (read) 19,000-20,000 IOPS (write)

CPU
Clock speed and the amount of available processors becomes a consideration if you anticipate using MapReduce. It has also been noted that when running a MongoDB instance with the majority of the data in memory, clock speed can have a major impact on overall performance. If you are planning to use MapReduce or you're able to operate with a majority of your data in memory, consider a deployment strategy that includes a CPU with a high clock/bus speed to maximize your operations per second.

Replication
Replication provides high availability of your data if a node fails in your cluster. It should be standard to replicate with at least three nodes in any MongoDB deployment. The most common configuration for replication with three nodes is a 2x1 deployment — having two primary nodes in a single data center with a backup server in a secondary data center:

MongoDB Replication

Sharding
If you anticipate a large, active data set, you should deploy a sharded MongoDB deployment. Sharding allows you to partition a single data set across multiple nodes. You can allow MongoDB to automatically distribute the data across nodes in the cluster or you may elect to define a shard key and create range-based sharding for that key.

Sharding may also help write performance, so you can also elect to shard even if your data set is small but requires a high amount of updates or inserts. It's important to note that when you deploy a sharded set, MongoDB will require three (and only three) config server instances which are specialized Mongo runtimes to track the current shard configuration. Loss of one of these nodes will cause the cluster to go into a read-only mode (for the configuration only) and will require that all nodes be brought back online before any configuration changes can be made.

Write Safety Mode
There are several write safety modes that govern how MongoDB will handle the persistence of the data to disk. It is important to consider which mode best fits your needs for both data integrity and performance. The following write safety modes are available:

None – This mode provides a deferred writing strategy that is non-blocking. This will allow for high performance, however there is a small opportunity in the case of a node failing that data can be lost. There is also the possibility that data written to one node in a cluster will not be immediately available on all nodes in that cluster for read consistency. The 'None' strategy will also not provide any sort of protection in the case of network failures. That lack of protection makes this mode highly unreliable and should only be used when performance is a priority and data integrity is not a concern.

Normal – This is the default for MongoDB if you do not select any other mode. It provides a deferred writing strategy that is non-blocking. This will allow for high performance, however there is a small opportunity in the case of a node failing that data can be lost. There is also the possibility that data written to one node in a cluster will not be immediately available on all nodes in that cluster for read consistency.

Safe – This mode will block until MongoDB has acknowledged that it has received the write request but will not block until the write is actually performed. This provides a better level of data integrity and will ensure that read consistency is achieved within a cluster.

Journal Safe – Journals provide a recovery option for MongoDB. Using this mode will ensure that the data has been acknowledged and a Journal update has been performed before returning.

Fsync - This mode provides the highest level of data integrity and blocks until a physical write of the data has occurred. This comes with a degradation in performance and should be used only if data integrity is the primary concern for your application.

Testing the Deployment
Once you've determined your deployment strategy, test it with a data set similar to your production data. 10gen has several tools to help you with load testing your deployment, and the console has a tool named 'benchrun' which can execute operations from within a JavaScript test harness. These tools will return operation information as well as latency numbers for each of those operations. If you require more detailed information about the MongoDB instance, consider using the mongostat command or MongoDB Monitoring Service (MMS) to monitor your deployment during the testing.

Installation

When performing the installation of MongoDB, a few considerations can help create both a stable and performance-oriented solution. 10gen recommends the use CentOS (64-bit) as the base operating system if at all possible. If you try installing MongoDB on a 32-bit operating system, you might run into file size limits that cause issues, and if you feel the urge to install it on Windows, you'll see performance issues if virtual memory begins to be utilized by the OS to make up for a lack of RAM in your deployment. As a result, 32-bit operating systems and Windows operating systems should be avoided on MongoDB servers. SoftLayer provisions CentOS 6.X 64-bit operating systems by default on all of our MongoDB engineered server deployments.

When you've got CentOS 64-bit installed, you should also make the following changes to maximize your performance (all of which are included by default on all SoftLayer engineered servers):

Set SSD Read Ahead Defaults to 16 Blocks - SSD drives have excellent seek times allowing for shrinking the Read Ahead to 16 blocks. Spinning disks might require slight buffering so these have been set to 32 blocks.

noatime - Adding the noatime option eliminates the need for the system to make writes to the file system for files which are simply being read — or in other words: Faster file access and less disk wear.

Turn NUMA Off in BIOS - Linux, NUMA and MongoDB tend not to work well together. If you are running MongoDB on NUMA hardware, we recommend turning it off (running with an interleave memory policy). If you don't, problems will manifest in strange ways like massive slow downs for periods of time or high system CPU time.

Set ulimit - We have set the ulimit to 64000 for open files and 32000 for user processes to prevent failures due to a loss of available file handles or user processes.

Use ext4 - We have selected ext4 over ext3. We found ext3 to be very slow in allocating files (or removing them). Additionally, access within large files is poor with ext3.

One last tip on installation: Make the Journal and Data volumes be distinct physical volumes. If the Journal and Data directories reside on a single physical volume, flushes to the Journal will interrupt the access of data and provide spikes of high latency within your MongoDB deployment.

Operations

Once a MongoDB deployment has been promoted to production, there are a few recommendations for monitoring and optimizing performance. You should always have the MMS agent running on all MongoDB instances to help monitor the health and performance of your deployment. Additionally, this tool is also very useful if you have 10gen MongoDB Cloud Subscriptions because it provides useful debugging data for the 10gen team during support interactions. In addition to MMS, you can use the mongostat command (mentioned in the deployment section) to see runtime information about the performance of a MongoDB node. If either of these tools flags performance issues, sharding or indexing are first-line options to resolve them:

Indexes - Indexes should be created for a MongoDB deployment if monitoring tools indicate that field based queries are performing poorly. Always use indexes when you are querying data based on distinct fields to help boost performance.

Sharding - Sharding can be leveraged when the overall performance of the node is suffering because of a large operating data set. Be sure to shard before you get in the red; the system only splits chunks for sharding on insert or update so if you wait too long to shard you may have some uneven distribution for a period of time or forever depending on your data set and sharding key strategy.

I know it seems like we've covered a lot over the course of this blog post, but this list of best practices is far from exhaustive. If you want to learn more, the MongoDB forums are a great resource to connect with the rest of the MongoDB community and learn from their experiences, and the documentation on MongoDB's site is another phenomenal resource. The best people to talk to when it comes to questions about MongoDB are the folks at 10gen, so I also highly recommend taking advantage of MongoDB Cloud Subscriptions to get their direct support for your one-off questions and issues.

-Harold

August 21, 2012

High Performance Computing - GPU v. CPU

Sometimes, technical conversations can sound like people are just making up tech-sounding words and acronyms: "If you want HPC to handle Gigaflops of computational operations, you probably need to supplement your server's CPU and RAM with a GPU or two." It's like hearing a shady auto mechanic talk about replacing gaskets on double overhead flange valves or hearing Chris Farley (in Tommy Boy) explain that he was "just checking the specs on the endline for the rotary girder" ... You don't know exactly what they're talking about, but you're pretty sure they're lying.

When we talk about high performance computing (HPC), a natural tendency is to go straight into technical specifications and acronyms, but that makes the learning curve steeper for people who are trying to understand why a solution is better suited for certain types of workloads than technology they are already familiar with. With that in mind, I thought I'd share a quick explanation of graphics processing units (GPUs) in the context of central processing units (CPUs).

The first thing that usually confuses people about GPUs is the name: "Why do I need a graphics processing unit on a server? I don't need to render the visual textures from Crysis on my database server ... A GPU is not going to benefit me." It's true that you don't need cutting-edge graphics on your server, but a GPU's power isn't limited to "graphics" operations. The "graphics" part of the name reflects the original intention for kind of processing GPUs perform, but in the last ten years or so, developers and engineers have come to adapt the processing power for more general-purpose computing power.

GPUs were designed in a highly parallel structure that allows large blocks of data to be processed at one time — similar computations are being made on data at the same time (rather than in order). If you assigned the task of rendering a 3D environment to a CPU, it would slow to a crawl — it handles requests more linearly. Because GPUs are better at performing repetitive tasks on large blocks of data than CPUs, you start see the benefit of enlisting a GPU in a server environment.

The Folding@home project and bitcoin mining are two of the most visible distributed computing projects that GPUs are accelerating, and they're perfect examples of workloads made exponentially faster with the parallel processing power of graphics processing units. You don't need to be folding protein or completing a blockchain to get the performance benefits, though; if you are taxing your CPUs with repetitive compute tasks, a GPU could make your life a lot easier.

If that still doesn't make sense, I'll turn the floor over to the Mythbusters in a presentation for our friends at NVIDIA:

SoftLayer uses NVIDIA Tesla GPUs in our high performance computing servers, so developers can use "Compute Unified Device Architecture" (CUDA) to easily take advantage of their GPU's capabilities.

Hopefully, this quick rundown is helpful in demystifying the "technobabble" about GPUs and HPC ... As a quick test, see if this sentence makes more sense now than it did when you started this blog: "If you want HPC to handle Gigaflops of computational operations, you probably need to supplement your server's CPU and RAM with a GPU or two."

-Phil

July 4, 2012

Cedexis: Tech Partner Spotlight

This guest blog features Cedexis, a featured member of the SoftLayer Technology Partners Marketplace. Cedexis a content and application delivery system that offers strategies and solutions for multi-platform content and application delivery to companies focused on maximizing web performance. In this video we talk to Cedexis Co-Founder Julien Coulon.

Company Website: www.cedexis.com
Tech Partners Marketplace: http://www.softlayer.com/marketplace/cedexis

A Multi-Cloud Strategy - The Key to Expansion and Conversion

Web and mobile applications have collapsed geographic barriers to business, bringing brand and commerce experiences ever-closer to increasingly far-flung customers. While web-based business models are powerful enablers for global expansion, they also create new a new challenge in managing availability and performance across diverse and distributed markets: How do you ensure consistent web performance across all markets without investing in physical infrastructure in all of those markets?

Once a business gets its core business on a consistent and reliable provider like SoftLayer, we typically recommend that they consider a multi-cloud strategy that will spread availability and performance risk across a global infrastructure of public and private data centers, delivery networks and cloud providers. Regardless of how fantastic your core SoftLayer hosting is, the reality is that single-source dependency introduces significant business risk. Fortunately, much of that business risk can be mitigated by adding a layer of multi-cloud architecture to support the application.

Recent high-profile outages speak to the problem that multi-sourcing solves, but many web-based operations remain precariously dependent on individual hosting, CDN and cloud providers. It's a lot like having server backups: If you never need a backup that you have, that backup probably isn't worth much to you, but if you need a backup that you don't have, you'd probably pay anything to have it.

A multi-cloud strategy drives revenue and other conversions. Why? Because revenue and conversions online correlate closely with a site's availability and performance. High Scalability posted several big-name real-world examples in the article, "Latency is Everywhere and it Costs You Sales." When an alternative vendor is just one click away, performance often makes a difference measured in dollars.

How Cedexis Can Help

Cedexis was founded to help businesses see and take advantage of a multi-cloud strategy when that strategy can provide better uptime, faster page loads, reliable transactions, and the ability to optimize cost across a diverse network of platforms and providers. We built the Cedexis Radar to measure the comparative performance of major cloud and delivery network providers (demo), and with that data, we created Openmix to provide adaptive automation for cloud infrastructure based on local user demand.

In order to do that effectively, Cedexis was built to be provider-agnostic, community-driven, actionable and adaptive. We support over 100 public cloud providers. We collect performance data based on crowd-sourced user requests (which represent over 900 million measurements per day from 32,000 individual networks). We allow organizations to write custom scripts that automate traffic routing based on fine-grained policies and thresholds. And we go beyond rules-driven traffic routing, dynamically matching actual user requests with the most optimal cloud at a specific moment in time.

Getting Started with Cedexis

  1. Join the Community
    Get real-time visibility into your users' performance.
  2. Compare the Performance of Your Clouds and Devliery Network
    Make informed decisions to optimize your site performance with Radar
  3. Leverage Openmix to optimize global web performance
    Optimize web and mobile performance to serve global markets

The more you can learn about your site, the more you can make it better. We want to help our customers drive revenue, enter new markets, avoid outages and reduce costs. As a SoftLayer customer, you've already found a fantastic hosting provider, and if Openmix won't provide a provable significant change, we won't sell you something you don't need. Our simple goal is to make your life better, whether you're a geek or a suit.

-Julien Coulon, Cedexis

This guest blog series highlights companies in SoftLayer's Technology Partners Marketplace.
These Partners have built their businesses on the SoftLayer Platform, and we're excited for them to tell their stories. New Partners will be added to the Marketplace each month, so stay tuned for many more come.
May 31, 2012

The Few. The Proud. The Red Herring Top 100.

Last week, I had the privilege of attending Red Herring's Top 100 North America Tech Award ceremony in Santa Monica. If you're not familiar with Red Herring, it harkens back to the headier days of the of the dot-com era in the late 90's and early 00's. While the markets have fluctuated quite a bit in the last dozen years, the startup scene has survived, and the optimism of the dot-com boom is still alive and well, albeit via more focused entrepreneurs that intentionally practice cold hard pragmatism and have bootstrap mentalities.

Today, the Red Herring Top 100 still serves as a great barometer for identifying promising new companies and entrepreneurs. The publication's editors are quick to point out that they were among the first to recognize that companies such as Facebook, Twitter, Google, Yahoo, Skype, Salesforce.com, YouTube and eBay would change the way we live and work. That's the start to a pretty nice little "alumni" list if you ask me.

How does a company make the cut?

The Top 100 were judged on both quantitative and qualitative criteria, such as financial performance, technology innovation, quality of management, IP creation, CAGR, execution of strategy, and disruption in their respective industries.

Before the Top 100 are selected, each finalist has an opportunity to pitch their business model and share why they think they should be included. I heard one entrepreneur say, "I have over a million dollars invested from my family and friends, so this can't fail." These businesses may have started as simple ideas, but they're fueled by an entrepreneurial passion that have pushed them to become truly remarkable. Many of the finalists had already reached a certain level of success and were trying to build and scale-out their ideas — everything from new mobile apps, open source and storage offerings to cloud and big data optimized solutions.

While preparing a little bit of information for SoftLayer's presentation, I was pleasantly surprised to see that more than 20 finalists for Red Herring's Top 100 Americas Award were active SoftLayer customers!

10gen, AppFirst, Backupify, BrightRoll, Clickable, Cloudant, Cloudera, CVision Technologies, MedAvante, OPOWER, Optify, PageFreezer Software, Refinery29, richrelevance, RingRevenue, SAY Media, TagMan, VigLink and Zencoder

After the editors made the tough decisions to narrow down the finalists to the Top 100 winners, SoftLayer was honored and excited to join 10gen, Backupify, Cloudera, CVision Technologies, MedAvante, PageFreezer Software, RingRevenue, VigLink and Zencoder. At least 10% of the 2012 Red Herring Americas Top 100 companies are using SoftLayer.

Red Herring Americas Award

Early in my tenure at SoftLayer, a colleague told me, "We aren't looking to be the next big thing, we are looking to enable it." That's probably not going to stop us from throwing our hat in the ring to be considered for the Global 100 this fall, though.

-Andre

April 17, 2012

High Performance Computing for Everyone

This guest blog was submitted by Sumit Gupta, senior director of NVIDIA's Tesla High Performance Computing business.

The demand for greater levels of computational performance remains insatiable in the high performance computing (HPC) and technical computing industries, as researchers, geophysicists, biochemists, and financial quants continue to seek out and solve the world's most challenging computational problems.

However, access to high-powered HPC systems has been a constant problem. Researchers must compete for supercomputing time at popular open labs like Oak Ridge National Labs in Tennessee. And, small and medium-size businesses, even large companies, cannot afford to constantly build out larger computing infrastructures for their engineers.

Imagine the new discoveries that could happen if every researcher had access to an HPC system. Imagine how dramatically the quality and durability of products would improve if every engineer could simulate product designs 20, 50 or 100 more times.

This is where NVIDIA and SoftLayer come in. Together, we are bringing accessible and affordable HPC computing to a much broader universe of researchers, engineers and software developers from around the world.

GPUs: Accelerating Research

High-performance NVIDIA Tesla GPUs (graphics processing units) are quickly becoming the go-to solution for HPC users because of their ability to accelerate all types of commercial and scientific applications.

From the Beijing to Silicon Valley — and just about everywhere in between — GPUs are enabling breakthroughs and discoveries in biology, chemistry, genomics, geophysics, data analytics, finance, and many other fields. They are also driving computationally intensive applications, like data mining and numerical analysis, to much higher levels of performance — as much as 100x faster.

The GPU's "secret sauce" is its unique ability to provide power-efficient HPC performance while working in conjunction with a system's CPU. With this "hybrid architecture" approach, each processor is free to do what it does best: GPUs accelerate the parallel research application work, while CPUs process the sequential work.

The result is an often dramatic increase in application performance.

SoftLayer: Affordable, On-demand HPC for the Masses

Now, we're coupling GPUs with easy, real-time access to computing resources that don't break the bank. SoftLayer has created exactly that with a new GPU-accelerated hosted HPC solution. The service uses the same technology that powers some of the world's fastest HPC systems, including dual-processor Intel E5-2600 (Sandy Bridge) based servers with one or two NVIDIA Tesla M2090 GPUs:

NVIDIA Tesla

SoftLayer also offers an on-demand, consumption-based billing model that allows users to access HPC resources when and how they need to. And, because SoftLayer is managing the systems, users can keep their own IT costs in check.

You can get more system details and pricing information here: SoftLayer HPC Servers

I'm thrilled that we are able to bring the value of hybrid HPC computing to larger numbers of users. And, I can't wait to see the amazing engineering and scientific advances they'll achieve.

-Sumit Gupta, NVIDIA - Tesla

Subscribe to performance