We are going to purchase a Linx cluster from a hardware vendor, but we are not sure what pre-installnation services that we should nogotiate in advance. Anybody has good advice? For example, what CFD benchmark test the vendor should do for us? Who will be responsile for integrating the cluster with our existing network infrastructure?

What documentation we need from the vendor?

Thank you.

Tom

> Anybody has good advice?

Run memory tests for 24-48 hours (a good supplier should do this anyway).

Run at full cpu load for a few hours monitoring the temperatures with the air conditioning on and with it turned down/off.

> For example, what CFD benchmark test the vendor should do for us?

The NAS Parallel Benchmarks are still probably the most useful CFD tests for understanding the performance of your hardware for a range of typical CFD algorithms.

If you run a commercial CFD code then run their parallel benchmarks.

> Who will be responsile for integrating the cluster with our existing network infrastructure?

Whoever acts as your system administrator. Whatever acts as a front end for the cluster will simply be a normal computer on your internal network. An external supplier is not going to be familiar with how your company organises your computer network and so is unlikely to perform this task efficiently even if you are prepared to give them root access to your servers.

> What documentation we need from the vendor?

Assuming your system administrator will look after the cluster in future then they need to know what hardware was bought, what software was installed, how it was configured, what issues arose during installation and commissioning and the benchmark performance. If you do it yourself this is typically a few tens of sides of hand written notes plus all the manuals that come with the hardware and software.

Hello Tom Smith,

Just curious; have you already decided that you Will by a certain cluster or will you decide after the hw vendor has done some benchmarks? I assume the latter. In that case I agree with andy. You could also let two (or more) vendors run the same benchmarks and see which one comes up with the best solution in terms of price/performance(/service). I also think that running the intended code on the cluster as a benchmark would be most beneficial, if that is possible.

regs

I am doing HPC hardware benchmark for my company now. We gave a minimum requirement to hardware vendors and ask them to propose the most suitable system for our environment. Interestingly, all vendors give different configurations. We got Unix and Linux for OS. RISC, Intel and AMD cpu. A very good experience.

Counting on hardware vendors to propose the best configuration for your environment is like going to a car dealership and asking which car they recommend for you. You will get a car that runs (hopefully) but don't count on the right number of doors, color, or engine for your needs. Vendors do not know enough (even if they are honestly trying their best) to sell you the right system. You need to educate yourself before taking on a major hpc cluster.

Thank you all.

We contacted several Linux cluster vendors, including the leading companies. To my surprise, few of them really know how to benchmark the cluster for CFD applications.

Since we will use commerical software, the company will provide benchmark utilities, but I am not sure what results we will get from the testing? Should we compare the performance with a existing standard cluster?

My concern is that after one year or two, the cluster may be slower and slower, then we don't know what is wrong, so maybe it is better to benchmark it now as a refernce. Do you agree?

BTW, any good Linux cluster book for a CFD engineer? We used Linux workstation before but we have not used cluster yet.

How to access a cluster at office from home in the evening or weekend?

Best regards,

Tom

> To my surprise, few of them really know how to benchmark the cluster for
: CFD applications.

This is normal since they will rarely sell hardware for CFD use. What is more of a problem is that they are unlikely to understand what is cost effective for CFD since it is rather different to that for accessing large databases and such. It also varies with the type of CFD program used: an explicit time stepping supersonic code may run best with lots of cheap processors and cheap interconnects whereas a strongly implicit code may be forced to use fewer faster more expensive processors with faster more expensive interconnects. Also the choice of motherboard and processors can make big differences: the higher clock speed processors are almost never cost effective but most things that can increase the speed of accessing main memory are cost effective for real CFD simulations.

> Since we will use commerical software, the company will provide
: benchmark utilities, but I am not sure what results we will get from the
: testing? Should we compare the performance with a existing standard
: cluster?

You compare the performance for the clusters on offer. Your commercial CFD provider is likely to provide a range of numbers for previously tested machines and so this will act as a form of check.

> My concern is that after one year or two, the cluster may be slower and
: slower, then we don't know what is wrong, so maybe it is better to
: benchmark it now as a refernce. Do you agree?

Computers do not get slower and slower with age but continue to perform at the same speed until something fails. It would be an unusual failure which allowed the computer to continue functioning at a reduced speed after failure although some data transfer problems may fit in this category.

After you receive the cluster the first thing you should do is check that it peforms to specification. Run the benchmarks you used to make your purchasing decision and if there is something wrong send it back.

> BTW, any good Linux cluster book for a CFD engineer?

There is plenty of information on the internet. The problem is recognising what is useful and what is misleading. Given the subject I strongly doubt there is a worthwhile book but there are useful articles such as those from people purchasing and installing their own clusters.

> How to access a cluster at office from home in the evening or weekend?

In exactly the same way as any other computer. For example, ssh to the front end and then run whatever command you want.