HPL Benchmark performance degradation of kernel 2.6.24.3 vs 2.6.23.14
Hi,
I have a five node intel Q6600 quad core cluster and I benchmarked
it with open source open mpi software using fc8 and it's supplied kernels recompiled and that of kernel.org with kernel 2.6.23.14 and 2.6.24.3.
With GotoBlas v 1.24 and open mpi beta both cases (v 1.3a) for kernels 2.6.23.14 with web100 i get 158GFlops.
But when i recompile with web100 for kernel 2.6.24 / without web100 and having 6gig DDR2800MHz ram on each node i get only 28GFLOPS AND 22GFLOPS for 5 nodes whereas with or without web 100 for kernel 2.6.23.14 i get 156-8 GfLOPS. wITH OR WITHOUT web 100 i get for kernel 2.6.24.3 22- 28 Gflops for 5 nodes.!
Why is there a performance drop in kernel 2.6.24.3 All else hardware is the same!
For inter node communication i use three pci express gig eth cards ( 2 intel and one syskonnect ) per node and using nptcp of netpipe their performance of intel and syskonnect cards in both kernels measured point to point is 880MBPS approx for all three cards with measured using netpipe for tcp with kernel 2.6.24.3 and 2.6.23.14 . I am also using three switches gigabit with high bisection b/w for these eth cards (copper) with 3 different subnets Yet I am getting a substantial performance drop keeping the hardware and openmpi and hpl and gotoblas same. Can some one help me figure out why?
Please find attached my kernel's .config
Cheers,
Allan Menezes
Quoted 9 lines · Allan Menezes
Hi,
I have a five node intel Q6600 quad core cluster and I benchmarked it
with open source open mpi software using fc8 and it's supplied kernels recompiled and that of kernel.org with kernel 2.6.23.14 and 2.6.24.3.
With GotoBlas v 1.24 and open mpi beta both cases (v 1.3a) for kernels 2.6.23.14 with web100 i get 158GFlops.
But when i recompile with web100 for kernel 2.6.24 / without web100 and having 6gig DDR2800MHz ram on each node i get only 28GFLOPS AND 22GFLOPS for 5 nodes whereas with or without web 100 for kernel 2.6.23.14 i get 156-8 GfLOPS. wITH OR WITHOUT web 100 i get for kernel 2.6.24.3 22- 28 Gflops for 5 nodes.!
Why is there a performance drop in kernel 2.6.24.3 All else hardware is the same!
For inter node communication i use three pci express gig eth cards ( 2 intel and one syskonnect ) per node and using nptcp of netpipe their performance of intel and syskonnect cards in both kernels measured point to point is 880MBPS approx for all three cards with measured using netpipe for tcp with kernel 2.6.24.3 and 2.6.23.14 . I am also using three switches gigabit with high bisection b/w for these eth cards (copper) with 3 different subnets Yet I am getting a substantial performance drop keeping the hardware and openmpi and hpl and gotoblas same. Can some one help me figure out why?
Please find attached my kernel's .config
Hi Allan
Your setup is quite complex, so you should give more information if you want some help here.
Is this benchmark stressing disk IO, task scheduler, network stack, memory, swap... hard to tell in fact.
Examining your .config, I would point out CONFIG_SLUB_DEBUG=y You really should disable this expensive option. (and possibly use CONFIG_SLAB instead of CONFIG_SLUB)
You probably should try to use oprofile tool, because its results are probably a good way to give hints about bad configuration, or kernel regressions.
opcontrol --vmfile=/boot/vmlinux-2.6.24.3 --start
<benchmarking>
opreport -l /boot/vmlinux-2.6.24.3
Quoted 2 lines · Eric Dumazet
Examining your .config, I would point out
CONFIG_SLUB_DEBUG=yYou really should disable this expensive option.
CONFIG_SLUB_DEBUG_ON is the expensive one.
Quoted 1 line
(and possibly use
CONFIG_SLABinstead ofCONFIG_SLUB)
That is a good thing to test indeed.
Quoted 21 lines · Eric Dumazet
Quoted 9 lines · Allan Menezes
Hi,
I have a five node intel Q6600 quad core cluster and I benchmarked
it with open source open mpi software using fc8 and it's supplied kernels recompiled and that of kernel.org with kernel 2.6.23.14 and 2.6.24.3.
With GotoBlas v 1.24 and open mpi beta both cases (v 1.3a) for kernels 2.6.23.14 with web100 i get 158GFlops.
But when i recompile with web100 for kernel 2.6.24 / without web100 and having 6gig DDR2800MHz ram on each node i get only 28GFLOPS AND 22GFLOPS for 5 nodes whereas with or without web 100 for kernel 2.6.23.14 i get 156-8 GfLOPS. wITH OR WITHOUT web 100 i get for kernel 2.6.24.3 22- 28 Gflops for 5 nodes.!
Why is there a performance drop in kernel 2.6.24.3 All else hardware is the same!
For inter node communication i use three pci express gig eth cards ( 2 intel and one syskonnect ) per node and using nptcp of netpipe their performance of intel and syskonnect cards in both kernels measured point to point is 880MBPS approx for all three cards with measured using netpipe for tcp with kernel 2.6.24.3 and 2.6.23.14 . I am also using three switches gigabit with high bisection b/w for these eth cards (copper) with 3 different subnets Yet I am getting a substantial performance drop keeping the hardware and openmpi and hpl and gotoblas same. Can some one help me figure out why?
Please find attached my kernel's .configHi Allan
Your setup is quite complex, so you should give more information if you want some help here.
Is this benchmark stressing disk IO, task scheduler, network stack, memory, swap... hard to tell in fact.
Examining your .config, I would point out
CONFIG_SLUB_DEBUG=yYou really should disable this expensive option. (and possibly useCONFIG_SLABinstead ofCONFIG_SLUB)
Quoted 5 lines
You probably should try to use oprofile tool, because its results are probably a good way to give hints about bad configuration, or kernel regressions.
opcontrol --vmfile=/boot/vmlinux-2.6.24.3 --start
<benchmarking>
opreport -l/boot/vmlinux-2.6.24.3
I am not the original reporter, to get good numbers HPL tests cpu and mostly networking speed (if more than one machine is being used), if local it test whichever interprocess communication is being used.
It is floating point with communications to sync the different processes together.
Generally if it is abnormally slow, you either have a errant process on a machine, a problem with one machine, or a problem with networking latency, or possibly a problem with some other latency.
I have never seen the scheduler make a big difference (unless the scheduler is really really broken), and it if configured for speed it does little or no swap (but I have seem machines that were tuned to page out early cause slight slow downs in the numbers when things should have nicely fit in memory), and it does little or no disk IO in the timed speed calculation areas.
It is pretty much all network latency and floating point speed.
Roger
Roger