Alexei Popov
alexei.a.popov@gmail.com
it seems you are using the benchmark that is simply
too small for your cluster. If you look here,
you'll find that it takes ca. 400-500 seconds to complete on 8 cores.
Actually, with the latest processors Firefly seems to need the
updated set of benchmarks, at least for parallel runs.
I'd suggest you to test performance and scalability of your
particular cluster/setup using job that takes at least 1-2 hours
to complete on 8 cores running on single node.
Your nodes are fast, and my guess is you are using 1 Gbit Ethernet?
This is most likely optimal solution for such a small cluster,
at least the price/performance ratio is very reasonable.
However, you need high quality Ethernet adapters and really
good switch. Formally you need 8-port switch but I doubt
if a typical 8-port switch will be good enough for your purposes -
so you can try to experiment with 16 or 24-port models.
regards,
Alexei
On Tue Mar 30 '10 1:45pm, Vyacheslav wrote
------------------------------------------
>Hi!
>Alex, we used your options for tests without appreciable improvement. We have destroyed our cluster, reinstall OS on all nodes and have again created the cluster – results are the same…
>Below I apply some results for MP2 input, taken from Performance page (in %):
>Cores xdlb=t bind=t blkdlb=0 nbiolb=1 sndbuf=256_rcvbuf=256
>8 100 100 100 100 100
>16 166.7 178.7 165.4 169.6 176.3
>24 127.4 121.8 121.8 127.6 124.1
>32 182.4 165.4 181.1 186.1 194.2
>40 194.2
>Options p2p=.t. xdlb=.t. are presented in all inputs. Influence of sndbuf and rcvbuf values variation from 64 up to 1024 is insignificant.
>Have you any ideas? We haven't. Unfortunately, we have no any IT expert - both of us only chemists.
>Sorry, I've overlooked to mention, our HPC version is intended for Server 2008 Standard, not for Server 2008 R2.
>------------------------------------------------------------------------
>On Sat Mar 27 '10 10:08pm, Alex Granovsky wrote
>-----------------------------------------------
>>Hi,
>>sorry for the delayed reply.
>>It your HPC solution is based on Win2K8 R2, (note not R1),
>>you may be affected by the same core parking problem, as Windows 7 users.
> Note we do not test it under Win2K8 R2...
>>Second, just adding
>>
$p2p p2p=.t. $end
>>does not automatically enables dynamic load balancing. Use at least
>>
$p2p p2p=.t. dlb=.t. $end
>>or
>>
$p2p p2p=.t. xdlb=.t. $end
>>You can also give a try to more advanced communication-related p2p options like:
>>
$p2p p2p=.t. xdlb=.t. ! to bind to interface bind=.t. ! to use non-blocked I/O for DLB blkdlb=0 ! to use polling instead of blocking waiting for DLB event nbiolb=1 ! to set larger TCP window sizes (in Kbytes) ! for send and receive buffers. sndbuf=256 rcvbuf=256 $end
>>Finally, use "-daf 2" command line option to reduce communication's
>>overhead.
>>regards,
>>Alex Granovsky
>
>
>