Hack bench regression with CONFIG_SLUB_CPU_PARTIAL disabled (info only)
Subsystems (2)
Hi,
While testing hackbench with 3.11-rc1 I observed hackbench testcase took too long to complete (32 vcpu kvm guest on 32 core HT off machine).
hackbench 1x degraded by 3 time (40sec vs 112sec) hackbench 2x degraded by around 20time (90sec vs 1800 sec)
When I bisected, I found that I had CONFIG_SLUB_CPU_PARTIAL
accidentally disabled.
I understand that CONFIG_SLUB_CPU_PARTIAL=y by default and the config
option is to provide determinism for RT kernels and thus it is not a
problem.
But thought of sharing this here so that if somebody runs into same issue accidently, it would save their time.
Quoted 8 lines · Raghavendra K T
Hi,
While testing hackbench with 3.11-rc1 I observed hackbench testcase took too long to complete (32 vcpu kvm guest on 32 core HT off machine).
hackbench 1x degraded by 3 time (40sec vs 112sec) hackbench 2x degraded by around 20time (90sec vs 1800 sec)
Quoted 7 lines
When I bisected, I found that I had
CONFIG_SLUB_CPU_PARTIALaccidentally disabled.
I understand thatCONFIG_SLUB_CPU_PARTIAL=yby default and the config option is to provide determinism for RT kernels and thus it is not a problem.
But thought of sharing this here so that if somebody runs into same issue accidently, it would save their time.
Incidentally, what difference do you currently get with SLUB and SLAB ?
Quoted 17 lines · Eric Dumazet
Quoted 8 lines · Raghavendra K T
Hi,
While testing hackbench with 3.11-rc1 I observed hackbench testcase took too long to complete (32 vcpu kvm guest on 32 core HT off machine).
hackbench 1x degraded by 3 time (40sec vs 112sec) hackbench 2x degraded by around 20time (90sec vs 1800 sec)
Quoted 7 lines
When I bisected, I found that I had
CONFIG_SLUB_CPU_PARTIALaccidentally disabled.
I understand thatCONFIG_SLUB_CPU_PARTIAL=yby default and the config option is to provide determinism for RT kernels and thus it is not a problem.
But thought of sharing this here so that if somebody runs into same issue accidently, it would save their time.
Quoted 2 lines
Incidentally, what difference do you currently get with SLUB and SLAB ?
tested now,
hackbench 1x is almost 2 times slower (40 vs 93)
hackbench 2x is almost 6times slower (90sec vs 528)