How to measure the effect of huge pages ?
|[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]|
Hi,This would be more appropriate to a linux-help mailing list (which doesn't exist), as I think I do understand the kernel issues involved, but I do not see the effect I expect to see.
A week ago I updated Debian Testing - one of the packages updated was the Linux kernel, which went from 2.6.32.n to 126.96.36.199:
$ uname -aLinux super 2.6.38-2-amd64 #1 SMP Thu Apr 7 04:28:07 UTC 2011 x86_64 GNU/Linux
Now, 2.6.38 has anonymous (transparent) huge page support: $ cat /proc/meminfo ... AnonHugePages: 2484224 kB So shortly (30 seconds) after rebooting, I did: $ echo always >/sys/kernel/mm/transparent_hugepage/enabled $ echo always >/sys/kernel/mm/transparent_hugepage/defrag which is still in effect: $ cat /sys/kernel/mm/transparent_hugepage/enabled [always] madvise never $ cat /sys/kernel/mm/transparent_hugepage/defrag [always] madvise never There's 4 Gbyte of RAM on this machine, and /proc/cpuinfo gives: processor : 0 vendor_id : GenuineIntel cpu family : 6 model : 15 model name : Intel(R) Core(TM)2 Quad CPU Q6600 @ 2.40GHz stepping : 11 cpu MHz : 2394.000 cache size : 4096 KB physical id : 0 siblings : 4 core id : 0 cpu cores : 4 apicid : 0 initial apicid : 0 fpu : yes fpu_exception : yes cpuid level : 10 wp : yesflags : fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx lm constant_tsc arch_perfmon pebs bts rep_good nopl aperfmperf pni dtes64 monitor ds_cpl vmx est tm2 ssse3 cx16 xtpr pdcm lahf_lm dts tpr_shadow vnmi flexpriority
bogomips : 4799.67 clflush size : 64 cache_alignment : 64 address sizes : 36 bits physical, 48 bits virtual power management: (4 times, for each core).This is the main application (which fills the machine 4 times / day, 16 hours / day):
$ ps uxww USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND ...hirlam 5528 0.0 0.0 56080 1492 ? S 17:35 0:01 mpirun --mca mpi_paffinity_alone 1 --mca mpi_yield_when_idle 1 -np 4 /scratch/hirlam/hl_home/MPI/lib/src/linuxgfortranmpi/bin/hlprog.x hirlam 5529 98.0 17.8 1090948 725416 ? R 17:35 99:42 /scratch/hirlam/hl_home/MPI/lib/src/linuxgfortranmpi/bin/hlprog.x hirlam 5530 99.1 16.8 1091748 683932 ? R 17:35 100:50 /scratch/hirlam/hl_home/MPI/lib/src/linuxgfortranmpi/bin/hlprog.x hirlam 5531 98.8 16.8 1086800 682432 ? R 17:35 100:34 /scratch/hirlam/hl_home/MPI/lib/src/linuxgfortranmpi/bin/hlprog.x hirlam 5532 98.5 16.9 1093752 686796 ? R 17:35 100:16 /scratch/hirlam/hl_home/MPI/lib/src/linuxgfortranmpi/bin/hlprog.x
...One would think such an application, which takes about 70 % of RAM would be a prime example of one that gets a speed-up from huge pages. However, the change in running time was unmeasurable.
Now, /proc/meminfo above shows that the huge pages *are* allocated, and the only reasonable way they are is that they are allocated to *this* application (I also see their allocation drop after the application finishes).
What do I have to do to determine why this doesn't have the desired effect ? Does (ordinary) malloc/free play a role ? What other system parameters can I study to get a handle on this ?
Thanks for any insight you can offer. -- Toon Moene - e-mail: toon@xxxxxxxxx - phone: +31 346 214290 Saturnushof 14, 3738 XG Maartensdijk, The Netherlands At home: http://moene.org/~toon/; weather: http://moene.org/~hirlam/ Progress of GNU Fortran: http://gcc.gnu.org/wiki/GFortran#news -- To unsubscribe from this list: send the line "unsubscribe linux-newbie" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.linux-learn.org/faqs