Re: [perfmon2] pfmon sampling issue about --follow-all
Status: Beta
Brought to you by:
seranian
From: Hucheng Z. <zho...@gm...> - 2008-12-04 08:47:41
|
Thank you, Eranian. Resolved. On Thu, Dec 4, 2008 at 1:41 AM, stephane eranian <er...@go...>wrote: > modprobe perfmon_pebs_core_smpl > > > On Wed, Dec 3, 2008 at 9:28 AM, Hucheng Zhou <zho...@gm...> > wrote: > > Thanks, Eranian. > > I have switched to root, that problem is gone. > > Another issue is about pebs mode, which seems is not supported, why? > > [root@osprey5 speccpu]# pfmon --follow-all --aggregate-results -e > > MEM_LOAD_RETIRED:L2_MISS --no-cmd-output --with-header --smpl-module=pebs > > --short-smpl-periods=20011 --long-smpl-periods=50021 --resolve-addresses > > --overflow-block --smpl-outfile=run-MEM_LOAD_RETIRED:L2_MISS-sampling -- > > /home/zhc/speccpu/bin/specinvoke -d > > /home/zhc/speccpu/benchspec/CINT2000/175.vpr/run/00000004 -e speccmds.err > -o > > speccmds.out -f speccmds.cmd > > cannot find requested kernel sampling format : pebs_core > > Thanks. > > On Wed, Dec 3, 2008 at 12:32 PM, stephane eranian < > er...@go...> > > wrote: > >> > >> On Wed, Dec 3, 2008 at 5:12 AM, Hucheng Zhou <zho...@gm...> > >> wrote: > >> > Note that I have run ulimit -s unlimited and ulimit -l unlimited. > >> > > >> > >> > >> You need to run as root to be able to do this otherwise I don't think > >> ulimit -l unlimited > >> works. You also need to do ulimit in the SAME shell as the one you run > >> pfmon from. > >> > >> > >> > >> > >> > >> > >> > >> > On Wed, Dec 3, 2008 at 12:09 PM, Hucheng Zhou <zho...@gm... > > > >> > wrote: > >> >> > >> >> Hi @ll: > >> >> I just hacked the runspec of SPEC to automatically run pfmon to > >> >> collect > >> >> counting information and sampling information. > >> >> However, I can correctly counting information in counting mode of > >> >> pfmon, > >> >> with --follow-all, but failed in sampling mode of pmon. > >> >> [zhc@osprey5 speccpu]$ pfmon --follow-all --aggregate-results -e > >> >> BUS_TRANS_BURST:ALL_AGENTS:BOTH_CORES --no-cmd-output --with-header > >> >> --short-smpl-periods=20011 --long-smpl-periods=50021 > >> >> --resolve-addresses > >> >> --overflow-block > >> >> --smpl-outfile=run-BUS_TRANS_BURST:ALL_AGENTS:BOTH_CORES-sampling -- > >> >> /home/zhc/speccpu/bin/specinvoke -d > >> >> /home/zhc/speccpu/benchspec/CINT2000/164.gzip/run/00000019 -e > >> >> speccmds.err > >> >> -o speccmds.out -f speccmds.cmd > >> >> unknown ptrace event 0 > >> >> Not enough memory to create perfmon context for [10126], > >> >> check your locked memory resource limit with limit or ulimit > >> >> session terminated by unrecoverable error > >> >> After remover the option --follow-all, pfmon can perform correctly, > >> >> but > >> >> the result sampling file is NULL. > >> >> [zhc@osprey5 speccpu]$ pfmon --aggregate-results -e > >> >> BUS_TRANS_BURST:ALL_AGENTS:BOTH_CORES --no-cmd-output --with-header > >> >> --short-smpl-periods=20011 --long-smpl-periods=50021 > >> >> --resolve-addresses > >> >> --overflow-block > >> >> --smpl-outfile=run-BUS_TRANS_BURST:ALL_AGENTS:BOTH_CORES-sampling -- > >> >> /home/zhc/speccpu/bin/specinvoke -d > >> >> /home/zhc/speccpu/benchspec/CINT2000/164.gzip/run/00000019 -e > >> >> speccmds.err > >> >> -o speccmds.out -f speccmds.cmd > >> >> Why I must use -follow-all is because that each SPEC benchmark is > >> >> triggered by /home/zhc/speccpu/bin/specinvoke, otherwise, pfmon just > >> >> can > >> >> collect the PMU information of specinvoke. > >> >> Thanks for the incoming suggestions. > >> >> > >> >> -- > >> >> Institute of High Performance Computing, > >> >> > >> >> Department of Computer Science and Technology, > >> >> > >> >> Tsinghua University, Beijing, China, 100084. > >> > > >> > > >> > > >> > -- > >> > Institute of High Performance Computing, > >> > > >> > Department of Computer Science and Technology, > >> > > >> > Tsinghua University, Beijing, China, 100084. > >> > > > > > > > > > -- > > Institute of High Performance Computing, > > > > Department of Computer Science and Technology, > > > > Tsinghua University, Beijing, China, 100084. > > > -- Institute of High Performance Computing, Department of Computer Science and Technology, Tsinghua University, Beijing, China, 100084. |