Comparing performance with sysbench: performance analysis


Sven Vermeulen Fri 19 April 2013

So in the past few posts I discussed how sysbench can be used to simulate some workloads, specific to a particular set of tasks. I used the benchmark application to look at the differences between the guest and host on my main laptop, and saw a major performance regression with the memory workload test. Let's view this again, using parameters more optimized to view the regressions:

$ sysbench --test=memory --memory-total-size=32M --memory-block-size=64 run
  Operations performed: 524288 (2988653.44 ops/sec)
  32.00 MB transferred (182.41 MB/sec)

  Operations performed: 524288 (24920.74 ops/sec)
  32.00 MB transferred (1.52 MB/sec)

$ sysbench --test=memory --memory-total-size=32M --memory-block-size=32M run
  Operations performed: 1 (  116.36 ops/sec)
  32.00 MB transferred (3723.36 MB/sec)

  Operations performed: 1 (   89.27 ops/sec)
  32.00 MB transferred (2856.77 MB/sec)

From looking at the code (gotta love Gentoo for making this obvious ;-) we know that the memory workload, with a single thread, does something like the following:

total_bytes = 0;
repeat until total_bytes >= memory-total-size:
  total_bytes += memory-block-size

  (start event timer)
  pointer -> buffer;
  while pointer <-> end-of(buffer)
    write somevalue at pointer
  (stop event timer)

Given that the regression is most noticeable when the memory-block-size is very small, the part of the code whose execution count is much different between the two runs is the mutex locking, global memory increment and the start/stop of event timer.

In a second phase, we also saw that mutex locking itself is not impacted. In the above case, we have 524288 executions. However, if we run the mutex workload this number of times, we see that this hardly has any effect:

$ sysbench --test=mutex --mutex-num=1 --mutex-locks=524288 --mutex-loops=0 run
Host:      total time:        0.0275s
Guest:     total time:        0.0286s

The code for the mutex workload, knowing that we run with one thread, looks like this:

mutex_locks = 524288
(start event timer)
  lock = get_mutex()
until mutex_locks = 0;
(stop event timer)

To check if the timer might be the culprit, let's look for a benchmark that mainly does timer checks. The cpu workload can be used, when we tell sysbench that the prime to check is 3 (as its internal loop runs from 3 till the given number, and when the given number is 3 it skips the loop completely) and we ask for 524288 executions.

$ sysbench --test=cpu --cpu-max-prime=3 --max-requests=524288 run
Host:  total time:  0.1640s
Guest: total time: 21.0306s

Gotcha! Now, the event timer (again from looking at the code) contains two parts: getting the current time (using clock_gettime()) and logging the start/stop (which is done in memory structures). Let's make a small test application that gets the current time (using the real-time clock as the sysbench application does) and see if we get similar results:

$ cat test.c
#include <stdio.h>
#include <time.h>

int main(int argc, char **argv, char **arge) {
  struct timespec tps;
  long int i = 524288;
  while (i-- > 0)
    clock_gettime(CLOCK_REALTIME, &tps);

$ gcc -lrt -o test test.c
$ time ./test
Host:  0m0.019s
Guest: 0m5.030s

So given that the clock_gettime() is ran twice in the sysbench, we already have 10 seconds of overhead on the guest (and only 0,04s on the host). When such time-related functions are slow, it is wise to take a look at the clock source configured on the system. On Linux, you can check this by looking at /sys/devices/system/clocksource/*.

# cd /sys/devices/system/clocksource/clocksource0
# cat current_clocksource
# cat available_clocksource
kvm-clock tsc hpet acpi_pm

Although kvm-clock is supposed to be the best clock source, let's switch to the tsc clock:

# echo tsc > current_clocksource

If we rerun our test application, we get a much more appreciative result:

$ time ./test
Host:  0m0.019s
Guest: 0m0.024s

So, what does that mean for our previous benchmark results?

$ sysbench --test=cpu --cpu-max-prime=20000 run
Host:            35,3049 sec
Guest (before):  36,5582 sec
Guest (now):     35,6416 sec

$ sysbench --test=fileio --file-total-size=6G --file-test-mode=rndrw --max-time=300 --max-requests=0 --file-extra-flags=direct run
Host:            1,8424 MB/sec
Guest (before):  1,5591 MB/sec
Guest (now):     1,5912 MB/sec

$ sysbench --test=memory --memory-block-size=1M --memory-total-size=10G run
Host:            3959,78 MB/sec
Guest (before)   3079,29 MB/sec
Guest (now):     3821,89 MB/sec

$ sysbench --test=threads --num-threads=128 --max-time=10s run
Host:            9765 executions
Guest (before):   512 executions
Guest (now):      529 executions

So we notice that this small change has nice effects on some of the tests. The CPU benchmark improves from 3,55% overhead to 0,95%; fileio is the same (from 15,38% to 13,63%), memory improves from 22,24% overhead to 3,48% and threads remains about status quo (from 94,76% slower to 94,58%).

That doesn't mean that the VM is now suddenly faster or better than before - what we changed was how fast a certain time measurement takes, which the benchmark software itself uses rigorously. This goes to show how important it is to

  1. understand fully how the benchmark software works and measures
  2. realize the importance of access to source code is not to be misunderstood
  3. know that performance benchmarks give figures, but do not tell you how your users will experience the system

That's it for the sysbench benchmark for now (the MySQL part will need to wait until a later stage).