Re: [RFC kvm-unit-tests] api: Add dirty logging performance test

[Takuya Yoshikawa <takuya.yoshikawa@gmail.com>]

On Sun, 08 Jan 2012 16:21:08 +0200
Avi Kivity <avi@redhat.com> wrote:

> On 01/08/2012 03:58 PM, Takuya Yoshikawa wrote:
> > Check how long it takes to get dirty log according to the number of
> > dirty pages, like:
> >
> >   get dirty log:     49 us for          1 dirty pages
> >   get dirty log:     49 us for          2 dirty pages
> >   get dirty log:     45 us for          4 dirty pages
> >   get dirty log:     41 us for          8 dirty pages
> >   get dirty log:     40 us for         16 dirty pages
> >   get dirty log:     44 us for         32 dirty pages
> >   get dirty log:     39 us for         64 dirty pages
> >   get dirty log:     42 us for        128 dirty pages
> >   get dirty log:     45 us for        256 dirty pages
> >   get dirty log:     53 us for        512 dirty pages
> >   get dirty log:     72 us for       1024 dirty pages
> >   get dirty log:     99 us for       2048 dirty pages
> >   get dirty log:    132 us for       4096 dirty pages
> >   get dirty log:    224 us for       8192 dirty pages
> >   get dirty log:    383 us for      16384 dirty pages
> >   get dirty log:    725 us for      32768 dirty pages
> >   get dirty log:   1412 us for      65536 dirty pages
> >   get dirty log:   2746 us for     131072 dirty pages
> >   get dirty log:   5455 us for     262144 dirty pages
> 
> Nice!

I forgot to add warming up, letting the guest scan the memory,
before starting the test.

So the number of shadow pages was increasing during the test.

> 
> > +
> > +void write_mem(volatile bool& running, volatile int& nr_dirty_pages,
> > +               void* logged_slot_virt)
> > +{
> > +    while (nr_dirty_pages >= 0) {
> > +        char* var = static_cast<char*>(logged_slot_virt);
> > +
> > +        while (!running) {
> > +            delay_loop(1000);
> > +        }
> > +        for (int i = 0; i < nr_dirty_pages; ++i) {
> > +            ++(*var);
> > +            var += page_size;
> > +        }
> > +        running = false;
> 
> You use running both to start this loop, and signal its end.  Better to
> use two variables.
> 
> But why use threads at all?  Just call this before reading the dirty
> log, no need for synchronization.

I re-used your dirty-log code a lot, will update.

> 
> > +    }
> > +}
> > +
> > +void check_dirty_log(mem_slot& slot,
> > +                     volatile bool& running,
> > +                     volatile int& nr_dirty_pages)
> > +{
> > +    slot.set_dirty_logging(true);
> > +    slot.update_dirty_log();
> > +
> > +    for (int i = 1; i <= nr_pages; i *= 2) {
> > +        struct timeval start_time, end_time;
> > +        long time_usec;
> > +
> > +        nr_dirty_pages = i;
> > +        running = true;
> > +        // wait until the guest finishes writing
> > +        while (running) {
> > +            delay_loop(1000);
> > +        }
> > +
> > +        gettimeofday(&start_time, NULL);
> > +        slot.update_dirty_log();
> > +        gettimeofday(&end_time, NULL);
> 
> Nicer to have a function that returns time in nanoseconds.

I don't know such an API which can be used in userspace.

If we can use nano-precision timer, I want to check other things
too, e.g. emulation.

	Takuya

-- 
Takuya Yoshikawa <takuya.yoshikawa@gmail.com>