[pucidbu]

Acasas,

I work on either a cluster computing with four nodes with 2 x 8-cores Intel Xeon E5-2650v2 2,6GHz each and on PCs with i7-3770 and i7-4790, and I can tell confirm you that already a single i7-3770 (a little old in this moment) is MUCH faster than two aforementioned xeons... when comparing them with the 4790, the ratio is almost 32 xeon cores to 4 i7-4790 cores, for Matlab, Ansys HFSS and CST. Maybe the xeon processors are made for servers, but for computing they just suck! The features shown in comparison on several sites (for example, http://www.velocitymicro.com/blog/xe...ts-difference/) are all craps.

People, if your scope is computing, don't buy XEON processors!!!!

[acasas]

Originally Posted by pucidbu

Acasas,

I work on either a cluster computing with four nodes with 2 x 8-cores Intel Xeon E5-2650v2 2,6GHz each and on PCs with i7-3770 and i7-4790, and I can tell confirm you that already a single i7-3770 (a little old in this moment) is MUCH faster than two aforementioned xeons... when comparing them with the 4790, the ratio is almost 32 xeon cores to 4 i7-4790 cores, for Matlab, Ansys HFSS and CST. Maybe the xeon processors are made for servers, but for computing they just suck! The features shown in comparison on several sites (for example, http://www.velocitymicro.com/blog/xe...ts-difference/) are all craps.

People, if your scope is computing, don't buy XEON processors!!!!

Thank´s Roberto, I do totally agree with you, XEON is not worth its price at all

[evcelica]

Originally Posted by pucidbu

Acasas,

I work on either a cluster computing with four nodes with 2 x 8-cores Intel Xeon E5-2650v2 2,6GHz each and on PCs with i7-3770 and i7-4790, and I can tell confirm you that already a single i7-3770 (a little old in this moment) is MUCH faster than two aforementioned xeons... when comparing them with the 4790, the ratio is almost 32 xeon cores to 4 i7-4790 cores, for Matlab, Ansys HFSS and CST. Maybe the xeon processors are made for servers, but for computing they just suck! The features shown in comparison on several sites (for example, http://www.velocitymicro.com/blog/xe...ts-difference/) are all craps.

People, if your scope is computing, don't buy XEON processors!!!!

Are you sure they are configured properly? Because they should not be that slow. The most common mistake I see is people not configuring the Memory correctly. It must be balanced among the channels, all channels populated for each CPU with the exact same RAM and capacity.
You can check this by looking in the case, or by using a program like AIDA64.

[pucidbu]

Originally Posted by evcelica

Are you sure they are configured properly? Because they should not be that slow. The most common mistake I see is people not configuring the Memory correctly. It must be balanced among the channels, all channels populated for each CPU with the exact same RAM and capacity.
You can check this by looking in the case, or by using a program like AIDA64.

I don't know precisely, but I've got good reasons to believe it is configured properly since it's been configured directly from the company we purchased the cluster from. Anyway, can an improper memory configuration slow down performances so much? Since its cost is approximately 100.000€ I just think it was delivered optimally configured.

Does AIDA64 run under linux?

[evcelica]

I've seen many improperly configured computers straight from the "professionals".
I've even had the "professionals" and IT departments adamantly tell me there was no problem with the memory configuration, then we configure it properly and voila, double the performance. Yeah, your'e welcome!

I don't know if AIDA 64 runs under Linux.
Lets start out simple, how much RAM does the machine have? We may be able to tell from this number.

[pucidbu]

Originally Posted by evcelica

I've seen many improperly configured computers straight from the "professionals".
I've even had the "professionals" and IT departments adamantly tell me there was no problem with the memory configuration, then we configure it properly and voila, double the performance. Yeah, your'e welcome!

I don't know if AIDA 64 runs under Linux.
Lets start out simple, how much RAM does the machine have? We may be able to tell from this number.

hi Erik, first thanks for your support. The cluster is configured as follows:

there are 5 nodes, one master node and four nodes, say 1, 2, 3 and 4. The master node has only 32GB or RAM and a Xeon E5-2603v2, in fact we almost never use it.

Then, all of the four nodes have 128GB of RAM and 2 x 8-cores Intel Xeon E5-2650v2 2,6GHz each, for a total of 64 cores and half tera of RAM.

What me and my colleagues have seen is that launching exactly the SAME simulation (of different softwares, like Matlab and Ansys HFSS for example) on a WorkStation (WS) with an i7-4790 (4 cores) and on 32 xeon cores, it runs much faster on the former! Precisely, as regards Matlab, the comparison has been done considering 4 i7 cores on one side, and 16 xeon cores on the other (simply because the Matlab license installed on the cluster only allows us to exploit 16 cores at a time). As regards HFSS, the comparison is "4 i7 cores Vs 32 xeon cores" (since the HFSS' license we have allows us to use up to 32 cores simultaneously), and the time taken to carry out the simulation is pretty much the same (but please, note the number of cores involved in the comparison).

Thank you

[evcelica]

128 GB sounds like it would be configured properly (at least have the right number of DIMMS) Wow, That is surprising.
What MPI are you using?
Can you run the MPI Test and report your latency and bandwidth?
How well do these programs you are using scale?
Are you running a scalable problem?
How well do the individual nodes perform? This will tell you is its network based or node based. But those machines sound like very nice machines.

[pucidbu]

Originally Posted by evcelica

128 GB sounds like it would be configured properly (at least have the right number of DIMMS) Wow, That is surprising.
What MPI are you using?
Can you run the MPI Test and report your latency and bandwidth?
How well do these programs you are using scale?
Are you running a scalable problem?
How well do the individual nodes perform? This will tell you is its network based or node based. But those machines sound like very nice machines.

On the cluster I'm using Intel MPI, on the WSs IBM Platform. I don't know how to run an MPI test to test the latency bandwidth, but however, I can tell you the 5 cluster nodes are connected by an infiniband switch (40Gbit/s), whereas the WSs are on a common Gbit LAN.

Regardless of the possible scalability of the problem, the comparison has been done for different softwares without using the MPI at first (thus locally), and just then using it. For example, I've run the same Matlab simulation that uses Mosek (a tool for convex optimization) on either a WS (1 x i7-4790) and a single cluster node (2 x 8-cores Intel Xeon E5-2650v2 2,6GHz), so 4 i7 cores Vs 16 xeon cores... Consider that Mosek does not use MPI but it does use all the available cores of the local machine as soon as it starts parallelizeing computing (with this I mean that for the first iterative passes it uses just one core of the local machine, then the problem becomes parallelizable and so it starts using as many cores as it can for each node of the tree (fast Branch&Bound algorithm)). Overall, the single WS (4 i7 cores) is faster than a single cluster's node (16 xeon cores) in a variable way according to the different steps of the algorithm, but anyway from a maximum of 60% to a minimum of 0% faster.

[evcelica]

Well that is confusing. How well do the XEON nodes do with just 4 processes? Are they comparable to the i7? They should be slightly slower, but shouldn't be 1/4 the speed!
It could be that what you are doing doesn't scale well and you are running into too much overhead.
Try comparing 4 cores to 4 cores. This should tell us whether it is the computer or the scaling.
Also, is hyper-threading turned off? Because it should be.
What about general CPU benchmarks like Linpak? Or something else to measure general CPU performance, memory bandwidth, and latency?


(Sorry it takes me so long to reply sometimes, I don't check back often)