On Thu, Sep 27, 2012 at 3:44 PM, Shaun Thomas <sthomas@optionshouse.com> wrote:
> On 09/27/2012 04:39 PM, Scott Marlowe wrote:
>
>> Clarification that the two base machines were about the same price.
>> 48 opteron cores (2.2GHz) or 16 xeon cores at ~2.6GHz. It's been a
>> few years, I'm not gonna testify to the exact numbers in court.
>
>
> Same here. We got really good performance on Opteron "a few years ago" too.
> :)
>
> But some more anecdotes... with the 4x8 E7450 Dunnington, our performance
> was OK. With the 2x6x2 X5675 Nehalem, it was ridiculous. Half the cores,
> 2.5x the speed, so far as pgbench was concerned. On every workload, on every
> level of concurrency I tried. Like you said, the 7450 dropped off at higher
> concurrency, but the 5675 kept on trucking.
>
> That's why I qualified my statement about Intel CPUs as "lately." They
> really seem to have cleaned up their server architecture.
Yeah, Intel's made a lot of headway on multi-core architecture since
then. But the 5620 etc series of the time were still pretty meh at
high concurrency compared to the opteron. The latest ones, which I've
tested now (40 hyperthreaded cores i.e 80 virtual cores) are
definitely faster than the now 4 year old 48 core opterons. But at a
much higher cost for a pretty moderate (20 to 30%) increase in
performance. OTOH, they don't "break down" past 40 to 100 connections
any more, so that's the big improvement to me.
How the curve looks like heading to 60+ threads is mildly interesting,
but how the server performs as you go past it was what worried me
before. Now both architectures seem to behave much better in such
"overload" scenarios.