require info on Itanium processor

jai · September 24, 2009, 01:56

how Itanium processor differ from other processor for solving cfd problem?

if suppose my processor is intel xeon/AMD, i need 4 processor to solve the problem of 5 million elements..

if suppose my processor is intel Itanium i need only 2 processor to solve the problem of 5 million elements..

Is these correct? how many no of processor require to solve the same problem with Itanium processor?

ghorrocks · September 24, 2009, 06:45

If you are using a 64 bit operating system then the things which limit the size problem you can do is:
1) Memory - you need enough memory to fit it in. This can be either in one machine or distributed across a cluster, either is OK.
2) Patience - big problems take longer to run. Faster CPUs help but when you start approaching mesh convergence for most engineering simulations you are frequently looking at runs taking days to weeks to complete. Again clusters help here by speeding the solution time.

jai · September 25, 2009, 06:24

dear sir
for solving 10 million elements Ansys suguessted to have 16 processor and 64GB of RAM.
now i'm looking for a server and i came to know that deploying cluster computing is more convenient instead of buying a single server.
Is it true? i hope this true from your above statement.
if i'm going for cluster computing means i have to buy 2 server (each consists of 8 processors) and i've more doubt on this netwoking...
1. whether i've to install cfx in all machines or not...
Is i have to buy any other software for clustering and solving..
pls explain the cluster setup in details

ghorrocks · September 25, 2009, 06:42

The recommendation for ANSYS is for FEA structural analysis. This has very little in common with CFD so cannot be used as a basis for sizing your computing requirements.

Clusters have better performance than single server/multi CPU/multi-core systems. This is because you get better parallel speedup in distributed parallel than local parallel.

If you are looking at performance versus dollars then the best option is to get a cluster of the W55xx CPUs, with only one CPU per machine and only running about 2 processes per machine. This means half the CPU cores are sitting idle. As the cost of the machines is small compared to the cost of the licenses the best return on investment is to do this, even though it seems wasteful.

Regardless as to whether you do local parallel or distributed parallel the licensing cost is the same. There is a little more setup in distributed parallel but it is well documented and support (and the forum!) can help you up and running.

jai · October 1, 2009, 02:58

Dear sir
i plan to buy 4 nodes, each consists of 8 processors and 2GB RAM per processor.
therefore my cluster has 32 processors and 64 GB of RAM.
so i plan to run only on 16 processors out of 32 processor to obtain best performance...
for these things, how many number of parallel licenses i have to buy?

if i'll get 16 licenses from ansys means, whether i've provision in CFX or windows server 2008 or HPC server to use only 4 processors out of 8 processors in a node ?

ghorrocks · October 1, 2009, 07:03

You will need 16 parallel licenses. You can tell each process which machine to go to so you can split the job up as you describe.

You are proposing quite a significant investment. I recommend you organise some trial licenses from ANSYS and some test hardware from your computer supplier so you can benchmark these systems for yourself. This way you can optimise the system before you spend cash - and also make sure you don't buy a lemon.

I was fortunate enough a few years ago to hire a system as a trial before I purchased it. It turned out the system only ran half as fast as expected because the system had a poor quality PCI controller chip on the motherboard. I would never have guessed it would have been a problem from the specifications or benchmarks. The trial saved me buying a lemon (4 of them in a mini-cluster actually). So for any reasonable sized purchase I recommend you benchmark the system before purchase.

September 24, 2009, 01:56	require info on Itanium processor	#1
jai Member jaikrishna Join Date: Mar 2009 Location: chennai Posts: 56 Rep Power: 17	how Itanium processor differ from other processor for solving cfd problem? if suppose my processor is intel xeon/AMD, i need 4 processor to solve the problem of 5 million elements.. if suppose my processor is intel Itanium i need only 2 processor to solve the problem of 5 million elements.. Is these correct? how many no of processor require to solve the same problem with Itanium processor?

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Superlinear speedup in OpenFOAM 13	msrinath80	OpenFOAM Running, Solving & CFD	18	March 3, 2015 05:36
CFX11 + Fortran compiler ?	Mohan	CFX	20	March 30, 2011 18:56
DecomposePar unequal number of shared faces	maka	OpenFOAM Pre-Processing	6	August 12, 2010 09:01
Unaligned accesses on IA64	andre	OpenFOAM	5	June 23, 2008 10:37
OpenFOAM with Cygwin	kitchener	OpenFOAM Installation	6	April 24, 2006 23:09

September 24, 2009, 06:45		#2
ghorrocks Super Moderator Glenn Horrocks Join Date: Mar 2009 Location: Sydney, Australia Posts: 17,703 Rep Power: 143	If you are using a 64 bit operating system then the things which limit the size problem you can do is: 1) Memory - you need enough memory to fit it in. This can be either in one machine or distributed across a cluster, either is OK. 2) Patience - big problems take longer to run. Faster CPUs help but when you start approaching mesh convergence for most engineering simulations you are frequently looking at runs taking days to weeks to complete. Again clusters help here by speeding the solution time.

September 25, 2009, 06:24		#3
jai Member jaikrishna Join Date: Mar 2009 Location: chennai Posts: 56 Rep Power: 17	dear sir for solving 10 million elements Ansys suguessted to have 16 processor and 64GB of RAM. now i'm looking for a server and i came to know that deploying cluster computing is more convenient instead of buying a single server. Is it true? i hope this true from your above statement. if i'm going for cluster computing means i have to buy 2 server (each consists of 8 processors) and i've more doubt on this netwoking... 1. whether i've to install cfx in all machines or not... Is i have to buy any other software for clustering and solving.. pls explain the cluster setup in details

September 25, 2009, 06:42		#4
ghorrocks Super Moderator Glenn Horrocks Join Date: Mar 2009 Location: Sydney, Australia Posts: 17,703 Rep Power: 143	The recommendation for ANSYS is for FEA structural analysis. This has very little in common with CFD so cannot be used as a basis for sizing your computing requirements. Clusters have better performance than single server/multi CPU/multi-core systems. This is because you get better parallel speedup in distributed parallel than local parallel. If you are looking at performance versus dollars then the best option is to get a cluster of the W55xx CPUs, with only one CPU per machine and only running about 2 processes per machine. This means half the CPU cores are sitting idle. As the cost of the machines is small compared to the cost of the licenses the best return on investment is to do this, even though it seems wasteful. Regardless as to whether you do local parallel or distributed parallel the licensing cost is the same. There is a little more setup in distributed parallel but it is well documented and support (and the forum!) can help you up and running.

October 1, 2009, 02:58		#5
jai Member jaikrishna Join Date: Mar 2009 Location: chennai Posts: 56 Rep Power: 17	Dear sir i plan to buy 4 nodes, each consists of 8 processors and 2GB RAM per processor. therefore my cluster has 32 processors and 64 GB of RAM. so i plan to run only on 16 processors out of 32 processor to obtain best performance... for these things, how many number of parallel licenses i have to buy? if i'll get 16 licenses from ansys means, whether i've provision in CFX or windows server 2008 or HPC server to use only 4 processors out of 8 processors in a node ?

October 1, 2009, 07:03		#6
ghorrocks Super Moderator Glenn Horrocks Join Date: Mar 2009 Location: Sydney, Australia Posts: 17,703 Rep Power: 143	You will need 16 parallel licenses. You can tell each process which machine to go to so you can split the job up as you describe. You are proposing quite a significant investment. I recommend you organise some trial licenses from ANSYS and some test hardware from your computer supplier so you can benchmark these systems for yourself. This way you can optimise the system before you spend cash - and also make sure you don't buy a lemon. I was fortunate enough a few years ago to hire a system as a trial before I purchased it. It turned out the system only ran half as fast as expected because the system had a poor quality PCI controller chip on the motherboard. I would never have guessed it would have been a problem from the specifications or benchmarks. The trial saved me buying a lemon (4 of them in a mini-cluster actually). So for any reasonable sized purchase I recommend you benchmark the system before purchase.