Reproduce command:
amplxe-cl -quiet -collect-with runsa -knob event-config=CPU_CLK_UNHALTED.THREAD,CORE_POWER.LVL0_TURBO_LICENSE,CORE_POWER.LVL1_TURBO_LICENSE,CORE_POWER.LVL2_TURBO_LICENSE -data-limit=0 -finalization-mode=none -result-dir="/export/users/linliu/vtune" -target-tmp-dir="/export/users/linliu/xxx" -no-summary -- ls
Vtune still writes temporary files to /tmp directory instead of the directory I specified. Temporary files of vtune on our benchmarks are too large to be stored in the root directory. As a result, it always throws error "no space left on device".
Anybody knows about this problem? Much appreciated!
ok.
Could you try to increase sampling interval? just add '-knob sampling-interval=10'
BTW, what is the size of trace for your benchmark on 5 secs collection with default and with sampling-interval options?
Kirill
連結已複製
Hi Lin.
Sorry for confusion. '-target-tmp-dir' option is used for remote collection case only.
Could you set TMPDIR directly and retry your collection
TMPDIR="/export/users/linliu/xxx" amplxe-cl -quiet -collect-with runsa -knob event-config=CPU_CLK_UNHALTED.THREAD,CORE_POWER.LVL0_TURBO_LICENSE,CORE_POWER.LVL1_TURBO_LICENSE,CORE_POWER.LVL2_TURBO_LICENSE -data-limit=0 -finalization-mode=none -result-dir="/export/users/linliu/vtune" -no-summary -- ls
Thanks, Kirill
KIRILL U. (Intel) wrote:Hi Lin.
Sorry for confusion. '-target-tmp-dir' option is used for remote collection case only.
Could you set TMPDIR directly and retry your collection
TMPDIR="/export/users/linliu/xxx" amplxe-cl -quiet -collect-with runsa -knob event-config=CPU_CLK_UNHALTED.THREAD,CORE_POWER.LVL0_TURBO_LICENSE,CORE_POWER.LVL1_TURBO_LICENSE,CORE_POWER.LVL2_TURBO_LICENSE -data-limit=0 -finalization-mode=none -result-dir="/export/users/linliu/vtune" -no-summary -- ls
Thanks, Kirill
Hi Kirill,
Thanks for your response. It works but I found even if I get 100G for my own directory, the space are all eaten by vtune. And the result directory takes over 80G. I wonder why it is such huge since I only collect 4 events. My benchmarks run like 15 minutes.
Best Wishes,
Lin
KIRILL U. (Intel) wrote:Hi Lin.
Could you tell me what is count of core on your system?
Could you share results directory for just 'ls' collection?
Does your benchmark utilize all cpus on system?
Thanks, Kirill
Hi Kirill,
I use Intel(R) Xeon(R) Platinum 8180 CPU @ 2.50GHz CPU with 56 physical cores. I am running the SPEC INT RATE benchmarks and all cpus are used.
The result directory of 'ls' collection is in attachments.
Best Wishes,
Lin
ok.
Could you try to increase sampling interval? just add '-knob sampling-interval=10'
BTW, what is the size of trace for your benchmark on 5 secs collection with default and with sampling-interval options?
Kirill
I tried and it reduced to 10G with -knob sampling-interval=10.
Do you mean -duration=5? I used -duration=5 and -knob sampling-interval=10 and the result directory is about 97M on our benchmarks.
Thank you again Kirill, you really helped a lot!
Best Wishes,
Lin
