Message-ID: <4F6C6C68.8040007@auul.pri.ee>
Date: Fri, 23 Mar 2012 14:28:24 +0200
From: Imre Oolberg <imre@auul.pri.ee>
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:5.0) Gecko/20110805 Icedove/5.0
MIME-Version: 1.0
To: pgsql-admin@postgresql.org
Subject: Re: deciding between amd and intel processor
References: <4F6915FD.804@auul.pri.ee>
	<CAOR=d=0nG_ZUOMPfcga2yfhyynRNA8HvLHSogZ7QcB4+t1G02g@mail.gmail.com>
In-Reply-To: 
 <CAOR=d=0nG_ZUOMPfcga2yfhyynRNA8HvLHSogZ7QcB4+t1G02g@mail.gmail.com>
Content-Type: text/plain; charset=ISO-8859-1; format=flowed
Content-Transfer-Encoding: quoted-printable

Hi!

Lots of thanks for your attention and comments! This time we went with=20
Dell AMD, with matching memory and processor speed of 1600 MHz, 6 x 600=20
gb 10k harddisks we intend to use in three mirrors striped together. We=20
didnt yet dare to include SSD in pure form or with H700 CacheCade.=20
(Although i read Gregg Smith's thougths about some ssd models which=20
should be in itself suitable for db.)


Imre

On 03/21/12 03:37, Scott Marlowe wrote:
> On Tue, Mar 20, 2012 at 5:42 PM, Imre Oolberg<imre@auul.pri.ee>  wrote:
>> Hi!
>>
>> My comerade has to decide for his so to say generic and a bit unpredic=
ted
>> PostgreSQL needs for web applications between Intel and AMD based serv=
er.
>> For now it has been coming down to these processor choices
>>
>> 1. Dell PowerEdge R810 which has 2 x Intel=AE Xeon=AE E7=AD-4830 2.13G=
Hz, 24M
>> cache 6.4 GT/s QPI, Turbo, HT, 8C, 1066MHz Max Mem processors
>
> Is this CPU really limited to 1066MHz memory?  That's positively
> ancient by today's standards.
>
>> as i see this makes total of 2 x 16 cores (or 2 x 16 threads) amd in t=
otal
>> of 2 x 24 MB L2 cache
>
> HT cores aren't 100% as fast as a real core.  They can get close
> depending on the application.  Generally the more CPU intensive the
> better they do.  The more you treat your CPUs like a data pump the
> less it matters.  It terms of performance, it's like something between
> 16 and 32 cores.  I'd guestimate it at about 24 or so.
>
>> 2. Dell PowerEdge R815 4 x AMD Opteron 6272, 2.1GHz, 16C, Turbo
>> CORE, 16M L2/16M L3, 1600Mhz Max Mem processors
>>
>> as i see this makes total of 4 x 16 cores and in total of 4 x 16 MB L2=
 cache
>
> Yep.  FAR more important is how many different memory banks you're now
> capable of throwing at the problem.  Each socket allows for a new
> memory controller on the machine, since the newer Intels and AMDs have
> integrated memory controllers.  This reduces contention between CPUs
> for memory access, and increases overall throughput.   Given the much
> slower main memory speed listed for the Intels I'd expect the AMD
> machine to stomp the Intel machine into the mud in terms of
> throughput.  If the Intels are REALLY 1600, then the AMDs would still
> get the nod, but it would be closer.
>
> So far based on what you've posted, I'd pick the AMD, mainly because
> it has more memory bandwidth.
>
>> Other computer components are quite similar
>>
>> - 128 GB 1333 dual ranked lv rdimm MHz memory for AMD
>
> Is the cost of 1600MHz memory that much more?  If it's a few hundred
> or something, get the 1600MHz memory.  If it's a few thousand, then
> yeah, it might be worth sticking to 1333MHz memory and buying more
> hard drives etc.
>
>> - 128 GB 1066 dual ranked lv rdimm MHz memory for Intel
>> - PERC H700 Integrated RAID Controller, 1GB NV Cache
>> - 4 x 600GB 10K RPM SAS 6Gbps 2.5in Hot=ADplug Hard Drive
>
> I'm not that familiar with the PERC H700.  While older model PERCs
> were at best of questionable performance, the newer ones apparently
> get decent reviews.  that's a pretty small drive set for a machine
> this massive.
>
>> Based on dell.com AMD variant comes about $14 k and Intel $17 k.
>
> Wow, that's a lot.  A similar machine with 5 drive bays, 5 300G SAS
> drives and an Areca 1880 SAS controller, and 4x16 AMD 6272 is right at
> $10k from aberdeeninc.com, and that's with a 5 year warranty.  For
> that extra $4k you could probably upgrade to something with 16 or more
> drives.  IO is king of db performance.  All the CPU in the world won't
> help if you're waiting on your drive subsystem.
>
>> And i got following questions in mind
>>
>> 1. do you suggest this Intel or this AMD configuration taking also int=
o
>> account price? while it seems obvious to go with AMD so to say price-w=
ise
>> are there some hidden rocks behind the surface going with AMD (like L2=
 cache
>> usage, overall system stability)?
>
> I'd recommend neither, as I'd rather chew off my own arm than ever
> have to deal with Dell again.  Way too many deals gone sour, and
> horribly unknowledgable sales staff for me to deal with anymore.  If I
> have a purchasing officer somewhere in a big company to deal with them
> they're OK.
>
> The L2 cache is no biggie.  both the AMD and Intel CPUs listed are
> pretty good performers.  The equivalent intel chips cost WAY more than
> the AMD ones tho, and honestly unless you're going for the top of the
> line fastest 10 core CPUs the Intels aren't gonna be much faster, if
> any.
>
>> 2. if it makes much sence to ask like this what could be considered on
>> normal postgresql workload (with web applications) reasonable balance
>> between cpu cores and memory system has?
>
> Without seeing some analysis of the current system, it's hard to say
> what you need to upgrade really.  What do tools like iostat, vmstat,
> sar, iotop, iftop and so on say about the current system under load?
>
>> 3. it is intended to use debian on this system, now squeeze and some d=
ay
>> wheezy, both hardwares most probably match quite well with debian and
>> postgresql?
>
> You want a pretty late model kernel for a 64 core machine.  Anything
> released in the last two years will likely be ok, so yeah, Debian
> Squeeze or Ubuntu 10.04 should be ok.
>
>> 4. it is intended not to use virtualization (KVM or Xen) but if needed=
, then
>> just use different pg clusters as in pg_lsclusters, and controlling
>> resources for each with shared_buffers etc, or should virtualization r=
ather
>> be considered?
>
> Only use virtualization if you really have to.  Most of the time
> you're better off with one cluster, or maybe two or three at most if
> the usage patterns are really different.  Virtulization has both
> performance implications as well as data security implications, due to
> questions about proper buffer flushing.
>
>> 5. i know that with four physical disks databases are generally recomm=
ended
>> to use raid10 , but what would be recommended raid setup with six driv=
es?
>
> RAID-10 on all 6 or RAID-1E on 5 with a hot spare is how I'd go.  But
> honestly, even 6 spinning disks seems like a low count.  Now 6 SSDs is
> another matter.  How big is his data set?  if it's pretty large then
> he needs spinners, and in that case, throwing more at the problem is a
> good idea.  A 2U case that can hold 8 3.5" drives can hold a TON of
> data, 8x3TB drives, even in RAID-10 gives you 12TBs.
>
> But all of this advice is kind of like the proverbial blind guy
> describing an elephant.  I'm not sure what the real workload looks
> like.  Yeah it's a web app, but that's kinda generic too.  Banking,
> content management, poker?  Your friend needs to profile his current
> system to see what's his choke points before he shells out $10k or
> more for a db server.  Be sad if all he needs is a RAID controller and
> a box of 16 hard drives to make his current system performant and he
> spent all that money on CPUs and memory he doesn't need, etc.
>