Re: LWLock cache line alignment

On Thu, Feb 03, 2005 at 06:26:16AM -0800, Simon Riggs wrote:
> > From: Tom Lane [mailto:tgl(at)sss(dot)pgh(dot)pa(dot)us] wrote
> > "Simon Riggs" <simon(at)2ndquadrant(dot)com> writes:
> > > It looks like padding out LWLock struct would ensure that
> > each of those
> > > were in separate cache lines?
> >
> > I've looked at this before and I think it's a nonstarter;
> > increasing the
> > size of a spinlock to 128 bytes is just not reasonable.
> > (Remember there
> > are two per buffer.)
>
> Well, the performance is unreasonably poor, so its time to do something,
> which might if it is unreasonable for the general case would need to be
> port specific.
>
> Also, there's no evidence it would actually help
> > anything, because the contention we have been able to measure
> > is on only
> > one particular lock (BufMgrLock) anyway. But feel free to
> > try it to see
> > if you can see a difference.
>
> Well, the Wierd Context Switching issue isn't normal contention, which I
> agree is stacked up against BufMgrLock.
>
> Overlapping cache lines doesn't cause measurable user space contention,
> just poor performance when the delay for cache-spoil, refetch is
> required many times more than the minimum (ideal).
>
> I'm thinking that the 128 byte cache line on Intel is sufficiently
> higher than the 64 byte cache line on AMD to tip us into different
> behaviour at runtime.
>

Would it be possible to "randomize" the placement of the LWLock
struct to reduce lock correlation in the buffers. i.e. Ensure that
buffer locks and allocated at least 1 cacheline apart. This may
allow the reduction in cache coherency and still maintain a compact
data structure. Maybe a simple mod X offset?

Ken