FDS Optimal Hardware Configuration

[Anthony Gauke]

Hello all,

I am a recent novice at the utilisation of FDS, and so far have primarily generated input files through Pyrosim.

The company where I work has the aim of expanding the amount of analysis done through the modelling of fire scenarios. Currently running FDS will take the best part of a day on most models, and occasionally larger projects have been run over the weekend for want of continual processing time. This is being run on a Dell Optiplex 3020 (i5-4590, 8GB RAM)

We are looking to invest in hardware dedicated to FDS computation. I have noted that such dedicated systems usually are clusters of computers.

Are there any guidelines as to the marginal efficiency around additional computers/cores/cache/RAM on dedicated FDS systems?

Our models have so far contained between 0.2 million and 3 million cells. Future projects could be much larger.

Thank you for any assistance,

Anthony

[Randy McDermott]

I'm not the expert in this, but since no one else is chiming in, I would aim for a small distributed memory Linux system.  I would maximize RAM and try to get some multiple of 4 cores per cpu.  For serial jobs with OpenMP you can get a factor of 2 speed up with 4 cores.  But generally, for larger jobs you want to distribute the workload with MPI, up to about 32^3 per mesh.  Getting all this to be "optimal" is not trivial.  You can search and see tons of discussion on this topic.  Perhaps my inadequate answer will spur others to point you in a better direction.

--
You received this message because you are subscribed to the Google Groups "FDS and Smokeview Discussions" group.
To unsubscribe from this group and stop receiving emails from it, send an email to fds-smv+u...@googlegroups.com.
To post to this group, send email to fds...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/fds-smv/1c0fdfc7-d1fc-4fcb-a602-fec92d1b1b16%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Glenn Forney]

I agree with randy - we have found that Linux is better suited than Windows for running FDS in parallel.  This requires however knowledgeable with Linux system administration to set it up and manage. various nodes on our cluster have 16GB, 32GB and 64GB of memory.  Each node has 2 CPUs with some having 4 cores per CPU and others having 6 cores per cpu.  All nodes share a common hard drive via NFS.  

Something else to consider is using a cloud service where you rent out computer time on clusters of computers.

To view this discussion on the web visit https://groups.google.com/d/msgid/fds-smv/CAC6wax3g6DK%2BkvoL9jTn_LQKg%2B%2B-yu9-ueHHW4ofEsKQoWDCxQ%40mail.gmail.com.

For more options, visit https://groups.google.com/d/optout.

--

Glenn Forney

[Anthony Gauke]

Thank you both for your responses.

Do you have any idea whether additional cache for each processor is necessary?

[Cian Davis]

A lot does depend on the size of models you want to run.

In my experience, the high end i7 chips have much better benchmark scores per core than Xeons. In addition, the cost per core (especially when you include RAM and motherboards) is lower for i7s.

On this basis, our cluster is built mainly around single-socket i7s, rather than Xeons. The models we run rarely exceed the capacity of a 4-core i7. We had to upgrade from 16GB of RAM to 32GB for a set of models that weighed in a 7.7m cells (off the top of my head, the model ended up using 24GB of RAM).

An item that is on my to-do list is to create some benchmark models and then distribute to the community to test various features. From there, we can get a better idea of what features help.

Regards,

Cian

On 7 July 2015 at 01:25, Anthony Gauke <gau...@gmail.com> wrote:

Thank you both for your responses.

Do you have any idea whether additional cache for each processor is necessary?

--

You received this message because you are subscribed to the Google Groups "FDS and Smokeview Discussions" group.
To unsubscribe from this group and stop receiving emails from it, send an email to fds-smv+u...@googlegroups.com.
To post to this group, send email to fds...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/fds-smv/afcebad4-80fc-41d4-8c37-45f36cc45d5e%40googlegroups.com.

[Róbert Kramli]

That is true that a single core performance of high end i7 can be better than Xeons.
(Currently I'm working with Intel Core i7 4790K)

The xeon's design princible was the stability, robustness, high availability. Designed for 7/24 servers, workstations with a lot of memory.

So if I can afford I will check for servers designed for HPC (high performance computing).

On my wish list:
1 pc Supermicro Super Server Barebone SYS-1028TR-T 1600 USD (This include 2 motherboard with 2 CPU socket on each, 1U rack chassis, 1000 W PSU)
up to 4 pcs Intel® Xeon® Processor E5-2667 v3 2100 USD (8 cores, 3,2GHz)
DDR4 ECC RAM - as needed (up to 2x512GB)

So it can start from 4500 USD (Barebone + 1 Xeon+RAM).
It can be upgraded up to 4 CPU (12000 USD) as needed.

I'm not sure that the total cost of this computer (including operation, maintenance cost) higher than the Core i7 computers.

Robert