Re: lazy vxid locks, v1

On 06/13/2011 02:29 PM, Kevin Grittner wrote:
> Stefan Kaltenbrunner wrote:
>
>> on that particular 40cores/80 threads box:
>
>> unpatched:
>
>> c40: tps = 107689.945323 (including connections establishing)
>> c80: tps = 101885.549081 (including connections establishing)
>
>> fast locks:
>
>> c40: tps = 215807.263233 (including connections establishing)
>> c80: tps = 118266.615543 (including connections establishing)
>
>> fast locks + lazy vxid:
>
>> c40: tps = 223308.779212 (including connections establishing)
>> c80: tps = 65738.046558 (including connections establishing)
>
> Is there any way to disable the HT (or whatever technology attempts
> to make each core look like 2)? In my benchmarking that has kept
> performance from tanking as badly when a large number of processes
> are contending for CPU.

I can do that tomorrow, but I have now done a fair amount of
benchmarking on that box using various tests and for CPU intense
workloads(various math stuff, parallel compiles of the linux kernel,
some inhouse stuff, and some other database) I usually get a 60-70x
speedup over just using a single core and most recent CPUs (this one is
actually a brand new Westmere-EX) showed pretty good scaling with
HT/threading.
I'm actually pretty sure that at leas in some benchmarks it was not HT
that was the real problem but rather our general inability to scale much
beyond 10-12 cores for reads and even worse for writes (due to WAL
contention).

Stefan