On the use of low thread high speed “gaming computers†to solve engineer simulations.

   Many of us in the linux community work only with software that is FOSS, which stands for Free open-source software. This is software that is not only open source but is available without licensing fees. There are many outstanding products out on the market today that are considered FOSS from the Firefox browser to most distributions of linux to the OpenOffice suite of text editing programs. However, there are times when FOSS is not an option, a good example is in my line of work supporting engineering software especially CAD tools and simulators. This software is not only costly but it is very restrictive. Each aspect of the software is charged. For example many of the simulators can run multithreaded. With one piece of software running on up to 16 threads for a single simulation.  More threads require more tokens, and we pay per token available. This puts us in a situation that we wish to maximize the amount we accomplish with as little threads as we can.

   If for example an engineer needs a simulation to finish to prove a design concept and it will take 6 hours to simulate at 1 thread he will want another token in order to use more threads. Using one token may buy him or her a reduction of 3 hours in simulation time, but the cost is that the tokens used for his or her simulation cannot be used by another engineer. The simple solution would be to keep on buying more and more tokens until every engineer has enough to run on the maximum number of threads at all times. If there are 5 engineers who run simulations that can run on 16 threads each for the cost of 5 tokens then we will need 25 tokens. Of course the simple solution rarely works. The cost of 25 is so high that it can easily bankrupt a medium sized company.

   Another solution would be to use less tokens but implement advance queuing software. This has the advantage that engineers can submit tasks and the servers running the simulations will run at all time (we hope) using the tokens we do have to the utmost. This strategy works well when deadlines are far away, but as they get close the fight for slots can grow.

  Since the limiting resource here is the number of threads we tried a different approach. As we are paying per thread we run, we should try to run threads as fast as possible (increasing our performance) rather then our throughput.  To further justify our reasoning we looked at creating benchmarks for our tools and comparing the amount of time it took to run a simulation compared to the number of threads we employed for it.

  The conclusion was:  Independent of software and the type of simulation we ran the performance increased exponentially to 4.5 threads and then leveled off. A shocking result given that the tools we used came out in different years and were produced by different venders.

   Given this information we concluded that if we ran 4 threads 25% faster on machine A (by overclocking) we could achieve better results then on machine B despite the same architecture.  This meant that for the near trivial price (compared to a server’s cost or additional tokens) of a modified desktop computer we could outperform a server with the maximum number of tokens we could purchase.

Our new system specifications:

   The total cost was approximately 1200 per unit after rebates. Assembly took about 3 hours. Overclocking was achieved at 4.7Ghz stable with the maximum recorded temperature at 70 C. The operating system is centos with the full desktop installed. The NICs have two connections link aggregated to our servers.

  To test overclocking we wrote a simple infinite loop floating point operation in perl and launched 8 instances of it while monitoring the results using a FOSS program called i7z. The hard drive only exists to provide a boot drive all other functions are performed via ssh and NFS exports. The units sit headless in our server room. It has been estimated that we have reduced overall simulation time across the company by 50% with only two units.

  The analogy we give is one of transportation. We have servers which function like buses. They can move a great deal of people at a time, which is great but buses are slow. Now we constructed high speed sports cars, these cars can only move a few people at a time but can move them much faster.

Isiah Schwartz

Teledyne Lecroy