This is the mail archive of the systemtap@sourceware.org mailing list for the systemtap project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

overhead testing results of systemtap/LKET


Hi, we've tested the overhead of systemtap/LKET with some benchmarks
on a ppc64 machine.

It shows the overhead of systemtap/LKET is acceptable generally.
But it will also cause significant overhead for some benchmark of
special behavior, e.g. dbench. Dbench calls kill() in a very high
frequency to check whether a task is complete, thus leads to a high
overhead.

We categorized the event hooks into five groups in the testing:
grp1 - syscall.entry, process
grp2 - syscall.return, process
grp3 - iosyscall, ioscheduler, scsi, aio, process
grp4 - tskdispatch, pagefault, netdev, process
grp5 - syscall.entry, syscall.return, process

All the results are
   (score1 - score2)/score2 * 100%,  where:
score1: the benchmark score when probed by systemtap
score2: the benchmark score without probing

dbench (<3% is noise)
--------------------
grp1            -14.4%
grp2            -33.1%
grp3            -7.92%
grp4            -13.6%
grp5            -43.3%

specjbb (<3% is noise)
---------------------
grp 1           -0.87%
grp 2           -0.67%
grp 4           +0.47%
grp 5           +0.05%

tiobench (<3% is noise)
----------------------
grp1       sequential reads      +1.45%
           sequential writes     -6.98%
           random reads          +0.57%
           random writes         -2.11%
grp2       sequential reads      +0.11%
           sequential writes     -5.81%
           random reads          +0.03%
           random writes         -2.11%
grp3       sequential reads      +1.42%
           sequential writes     -6.98%
           random reads          +0.51%
           random writes         -2.11%
grp4       sequential reads      +1.38%
           sequential writes     -5.81%
           random reads          +0.60%
           random writes         -2.11%
grp5       sequential reads      +0.22%
           sequential writes     -8.14%
           random reads          -0.10%
           random writes         -1.05%

Rawiobench (<3% is noise)
------------------------
grp1       sequential aioread()     0%
           sequential aiowrite()    0%
           random aioread()         0%
           random aiowrite()        0%
grp2       sequential aioread()     0%
           sequential aiowrite()    0%
           random aioread()         0%
           random aiowrite()        -0.82%
grp3       sequential aioread()     0%
           sequential aiowrite()    0%
           random aioread()         0%
           random aiowrite()        0%
grp4       sequential aioread()     0%
           sequential aiowrite()    0%
           random aioread()         +0.79%
           random aiowrite()        -0.82%
grp5       sequential aioread()     0%
           sequential aiowrite()    -6.41%
           random aioread()         +0.79%
           random aiowrite()        0%

Test environment:
Machine:  Open Power 720/ 8 cpus/ 2 cores/ 6GB RAM (tiobench use 1G)
Software: RHEL4-U3GA/ 2.6.17.2/ systemtap-20060718/ elfutils-0.122-0.4


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]