Re: hardware advice

On Thu, Sep 27, 2012 at 3:40 PM, Evgeny Shishkin <itparanoia(at)gmail(dot)com> wrote:
>
> On Sep 28, 2012, at 1:36 AM, Scott Marlowe <scott(dot)marlowe(at)gmail(dot)com> wrote:
>
>> On Thu, Sep 27, 2012 at 3:16 PM, Claudio Freire <klaussfreire(at)gmail(dot)com> wrote:
>>> On Thu, Sep 27, 2012 at 6:08 PM, David Boreham <david_list(at)boreham(dot)org> wrote:
>>>>>
>>>>> We went from Dunnington to Nehalem, and it was stunning how much better
>>>>> the X5675 was compared to the E7450. Sandy Bridge isn't quite that much of a
>>>>> jump though, so if you don't need that kind of bleeding-edge, you might be
>>>>> able to save some cash. This is especially true since the E5-2600 series has
>>>>> the same TDP profile and both use 32nm lithography.
>>>>
>>>> We use Opteron on a price/performance basis. Intel always seems to come up
>>>> with some way to make their low-cost processors useless (such as limiting
>>>> the amount of memory they can address).
>>>
>>> Careful with AMD, since many (I'm not sure about the latest ones)
>>> cannot saturate the memory bus when running single-threaded. So, great
>>> if you have a high concurrent workload, quite bad if you don't.
>>
>> Conversely, we often got MUCH better parallel performance from our
>> quad 12 core opteron servers than I could get on a dual 8 core xeon at
>> the time. The newest quad 10 core Intels are about as fast as the
>> quad 12 core opteron from 3 years ago. So for parallel operation, do
>> remember to look at the opteron. It was much cheaper to get highly
>> parallel operation on the opterons than the xeons at the time we got
>> the quad 12 core machine at my last job.
>
> But what about latency, not throughput?

It means little when you're building a server to handle literally
thousands of queries per seconds from hundreds of active connections.
The intel box would have simply fallen over under the load we were
handling on the 48 core opteron at the time. Note that under maximum
load we saw load factors in the 20 to 100 on that opteron box and
still got very good response times (average latency on most queries
was still in the single digits of milliseconds).

For single threaded or only a few threads, yeah, the intel was
slightly faster, but as soon as the real load of our web site hit the
machine it wasn't even close.