On Wed, Apr 15, 2015 at 2:55 AM, Robert Haas <
robertmhaas@gmail.com> wrote:
>
> On Wed, Apr 16, 2014 at 2:44 PM, Tom Lane <
tgl@sss.pgh.pa.us> wrote:
> > Merlin Moncure <
mmoncure@gmail.com> writes:
> >> Anyways, I'm still curious if you can post similar numbers basing the
> >> throttling on gross allocation counts instead of time. Meaning: some
> >> number of buffer allocations has to have occurred before you consider
> >> eviction. Besides being faster I think it's a better implementation:
> >> an intermittently loaded server will give more consistent behavior.
> >
> > Yeah --- I think wall-clock-based throttling is fundamentally the wrong
> > thing anyway. Are we going to start needing a CPU speed measurement to
> > tune the algorithm with? Not the place to be going. But driving it off
> > the number of allocations that've been done could be sensible. (OTOH,
> > that means you need a central counter, which itself would be a
> > bottleneck.)
>
> So, I was thinking about this a little bit more today, prodded by my
> coworker John Gorman. I'm wondering if we could drive this off of the
> clock sweep; that is, every time the clock sweep touches a buffer, its
> usage count goes down by one, but we also set two flag bits. Every
> time the buffer gets touched thereafter, we check whether any flag
> bits are set; if so, we clear one and increase the usage count, else
> we do nothing. So the usage count can increase at most twice per
> clock sweep. The advantage of that is that, as with Peter's approach,
> it is harder for the usage count of a buffer to max out - to get
> there, you need sustained access over a longer period of time. But
> it's not time-dependent, so replaying the same workload at twice the
> speed or half the speed on faster or slower hardware doesn't change
> the choice of which buffer to evict, which seems good. And it will
> cause us to prefer buffers which are accessed repeatedly over a period
> of time rather than buffers that are accessed a bunch of times in
> quick succession and then not touched again for a while, which seems
> like a good bet.
>
IIUC, this will allow us to increase usage count only when the buffer
is touched by clocksweep to decrement the usage count.
I think such a solution will be good for the cases when many evictions
needs to be performed to satisfy the workload, OTOH when there are
not too many evictions that needs to be done, in such a case some of
the buffers that are accessed much more will have equal probability to
get evicted as compare to buffers which are less accessed.