Обсуждение: Xeon Woodcrest/Dempsey vs Opteron Socket F/940 with postgresql and some SAS raid-figures

Hi,

We've been running our "webapp database"-benchmark again on mysql and
postgresql. This time using a Fujitsu-Siemens RX300 S3 machine equipped
with a 2.66Ghz Woodcrest (5150) and a 3.73Ghz Dempsey (5080). And
compared those results to our earlier undertaken Opteron benchmarks on
2.4GHz' Socket F- and 940-versions (2216, 280).

You can see the english translation here:
http://tweakers.net/reviews/646

The Woodcrest is quite a bit faster than the Opterons. Actually... With
Hyperthreading *enabled* the older Dempsey-processor is also faster than
the Opterons with PostgreSQL. But then again, it is the top-model
Dempsey and not a top-model Opteron so that isn't a clear win.
Of course its clear that even a top-Opteron wouldn't beat the Dempsey's
as easily as it would have beaten the older Xeon's before that.

Again PostgreSQL shows very good scalability, so good even
HyperThreading adds extra performance to it with 4 cores enabled...
while MySQL in every version we tested (5.1.9 is not displayed, but
showed similar performance) was slower with HT enabled.

Further more we received our ordered Dell MD1000 SAS-enclosure which has
15 SAS Fujitsu MAX3036RC disks and that unit is controlled using a Dell
PERC 5/e.
We've done some benchmarks (unfortunately everything is in Dutch for this).

We tested varying amounts of disks in RAID10 (a set of 4,5,6 and 7
2-disk-mirrors striped), RAID50 and RAID5. The interfaces to display the
results are in a google-stylee beta-state, but here is a list of all
benchmarks done:
http://tweakers.net/benchdb/search?query=md1000&ColcomboID=5

Hover over the left titles to see how many disks and in what raid-level
  was done. Here is a comparison of 14 disk RAID5/50/10's:

http://tweakers.net/benchdb/testcombo/wide/?TestcomboIDs%5B1156%5D=1&TestcomboIDs%5B1178%5D=1&TestcomboIDs%5B1176%5D=1&DB=Nieuws&Query=Keyword

For raid5 we have some graphs:
http://tweakers.net/benchdb/testcombo/1156
Scroll down to see how adding disks improves performance on it. The
Areca 1280 with WD Raptor's is a very good alternative (or even better)
as you can see for most benchmarks, but is beaten as soon as the
relative weight of random-IO increases (I/O-meter fileserver and
database benchmarks), the processor on the 1280 is faster than the one
on the Dell-controller so its faster in sequential IO.
These benchmarks were not done using postgresql, so you shouldn't read
them as absolute for all your situations ;-) But you can get a good
impression I think.

Best regards,

Arjen van der Meijden
Tweakers.net

Very nice!

The 3Ware cards have fallen far behind Areca it seems.  They look close in
Raid 10 performance, but with RAID5 they get crushed.

I'm about to purchase 20 machines for the lab and I think this article
pushes me toward Woodcrest, though I think it's a short term decision with
quad core AMD socket F coming later this year.  Right now it seems that the
Intel advantage is about 30%-40%.

- Luke


On 9/7/06 10:51 PM, "Arjen van der Meijden" <acmmailing@tweakers.net> wrote:

> Hi,
>
> We've been running our "webapp database"-benchmark again on mysql and
> postgresql. This time using a Fujitsu-Siemens RX300 S3 machine equipped
> with a 2.66Ghz Woodcrest (5150) and a 3.73Ghz Dempsey (5080). And
> compared those results to our earlier undertaken Opteron benchmarks on
> 2.4GHz' Socket F- and 940-versions (2216, 280).
>
> You can see the english translation here:
> http://tweakers.net/reviews/646
>
> The Woodcrest is quite a bit faster than the Opterons. Actually... With
> Hyperthreading *enabled* the older Dempsey-processor is also faster than
> the Opterons with PostgreSQL. But then again, it is the top-model
> Dempsey and not a top-model Opteron so that isn't a clear win.
> Of course its clear that even a top-Opteron wouldn't beat the Dempsey's
> as easily as it would have beaten the older Xeon's before that.
>
> Again PostgreSQL shows very good scalability, so good even
> HyperThreading adds extra performance to it with 4 cores enabled...
> while MySQL in every version we tested (5.1.9 is not displayed, but
> showed similar performance) was slower with HT enabled.
>
> Further more we received our ordered Dell MD1000 SAS-enclosure which has
> 15 SAS Fujitsu MAX3036RC disks and that unit is controlled using a Dell
> PERC 5/e.
> We've done some benchmarks (unfortunately everything is in Dutch for this).
>
> We tested varying amounts of disks in RAID10 (a set of 4,5,6 and 7
> 2-disk-mirrors striped), RAID50 and RAID5. The interfaces to display the
> results are in a google-stylee beta-state, but here is a list of all
> benchmarks done:
> http://tweakers.net/benchdb/search?query=md1000&ColcomboID=5
>
> Hover over the left titles to see how many disks and in what raid-level
>   was done. Here is a comparison of 14 disk RAID5/50/10's:
> http://tweakers.net/benchdb/testcombo/wide/?TestcomboIDs%5B1156%5D=1&Testcombo
> IDs%5B1178%5D=1&TestcomboIDs%5B1176%5D=1&DB=Nieuws&Query=Keyword
>
> For raid5 we have some graphs:
> http://tweakers.net/benchdb/testcombo/1156
> Scroll down to see how adding disks improves performance on it. The
> Areca 1280 with WD Raptor's is a very good alternative (or even better)
> as you can see for most benchmarks, but is beaten as soon as the
> relative weight of random-IO increases (I/O-meter fileserver and
> database benchmarks), the processor on the 1280 is faster than the one
> on the Dell-controller so its faster in sequential IO.
> These benchmarks were not done using postgresql, so you shouldn't read
> them as absolute for all your situations ;-) But you can get a good
> impression I think.
>
> Best regards,
>
> Arjen van der Meijden
> Tweakers.net
>
> ---------------------------(end of broadcast)---------------------------
> TIP 3: Have you checked our extensive FAQ?
>
>                http://www.postgresql.org/docs/faq
>

Hi, Arjen,


On 8-Sep-06, at 1:51 AM, Arjen van der Meijden wrote:

> Hi,
>
> We've been running our "webapp database"-benchmark again on mysql
> and postgresql. This time using a Fujitsu-Siemens RX300 S3 machine
> equipped with a 2.66Ghz Woodcrest (5150) and a 3.73Ghz Dempsey
> (5080). And compared those results to our earlier undertaken
> Opteron benchmarks on 2.4GHz' Socket F- and 940-versions (2216, 280).
>
> You can see the english translation here:
> http://tweakers.net/reviews/646
>
> The Woodcrest is quite a bit faster than the Opterons. Actually...
> With Hyperthreading *enabled* the older Dempsey-processor is also
> faster than the Opterons with PostgreSQL. But then again, it is the
> top-model Dempsey and not a top-model Opteron so that isn't a clear
> win.
> Of course its clear that even a top-Opteron wouldn't beat the
> Dempsey's as easily as it would have beaten the older Xeon's before
> that.

Why wouldn't you use a top of the line Opteron ?
>
> Again PostgreSQL shows very good scalability, so good even
> HyperThreading adds extra performance to it with 4 cores enabled...
> while MySQL in every version we tested (5.1.9 is not displayed, but
> showed similar performance) was slower with HT enabled.
>
> Further more we received our ordered Dell MD1000 SAS-enclosure
> which has 15 SAS Fujitsu MAX3036RC disks and that unit is
> controlled using a Dell PERC 5/e.
> We've done some benchmarks (unfortunately everything is in Dutch
> for this).
>
> We tested varying amounts of disks in RAID10 (a set of 4,5,6 and 7
> 2-disk-mirrors striped), RAID50 and RAID5. The interfaces to
> display the results are in a google-stylee beta-state, but here is
> a list of all benchmarks done:
> http://tweakers.net/benchdb/search?query=md1000&ColcomboID=5
>
> Hover over the left titles to see how many disks and in what raid-
> level  was done. Here is a comparison of 14 disk RAID5/50/10's:
> http://tweakers.net/benchdb/testcombo/wide/?TestcomboIDs%5B1156%
> 5D=1&TestcomboIDs%5B1178%5D=1&TestcomboIDs%5B1176%
> 5D=1&DB=Nieuws&Query=Keyword
>
> For raid5 we have some graphs:
> http://tweakers.net/benchdb/testcombo/1156
> Scroll down to see how adding disks improves performance on it. The
> Areca 1280 with WD Raptor's is a very good alternative (or even
> better) as you can see for most benchmarks, but is beaten as soon
> as the relative weight of random-IO increases (I/O-meter fileserver
> and database benchmarks), the processor on the 1280 is faster than
> the one on the Dell-controller so its faster in sequential IO.
> These benchmarks were not done using postgresql, so you shouldn't
> read them as absolute for all your situations ;-) But you can get a
> good impression I think.
>
> Best regards,
>
> Arjen van der Meijden
> Tweakers.net
>
> ---------------------------(end of
> broadcast)---------------------------
> TIP 3: Have you checked our extensive FAQ?
>
>               http://www.postgresql.org/docs/faq
>

Dave Cramer wrote:
> Hi, Arjen,
>
>
>> The Woodcrest is quite a bit faster than the Opterons. Actually...
>> With Hyperthreading *enabled* the older Dempsey-processor is also
>> faster than the Opterons with PostgreSQL. But then again, it is the
>> top-model Dempsey and not a top-model Opteron so that isn't a clear win.
>> Of course its clear that even a top-Opteron wouldn't beat the
>> Dempsey's as easily as it would have beaten the older Xeon's before that.
>
> Why wouldn't you use a top of the line Opteron ?

What do you mean by this question? Why we didn't test the Opteron 285
instead of the 280?

Well, its not that you can just go up to a hardware supplier and pick
exactly the system you want to review/benchmar... especially not with
pre-production hardware that (at the time) wasn't very widely available.
Normally, you just get what system they have available at their
marketing or pre-sales department.

The Opteron 280 was from an earlier review and was fitted in the "Try
and Buy"-version of the Sun Fire x4200. In that system; you only have a
few options where the 280 was the fastest at the time.

But then again, systems with the Woodcrest 5150 (the subtop one) and
Opteron 280 (also the subtop one) are about equal in price, so its not a
bad comparison in a bang-for-bucks point of view. The Dempsey was added
to show how both the Opteron and the newer Woodcrest would compete
against that one.

Best regards,

Arjen

On 8-Sep-06, at 8:44 AM, Arjen van der Meijden wrote:

> Dave Cramer wrote:
>> Hi, Arjen,
>>> The Woodcrest is quite a bit faster than the Opterons.
>>> Actually... With Hyperthreading *enabled* the older Dempsey-
>>> processor is also faster than the Opterons with PostgreSQL. But
>>> then again, it is the top-model Dempsey and not a top-model
>>> Opteron so that isn't a clear win.
>>> Of course its clear that even a top-Opteron wouldn't beat the
>>> Dempsey's as easily as it would have beaten the older Xeon's
>>> before that.
>> Why wouldn't you use a top of the line Opteron ?
>
> What do you mean by this question? Why we didn't test the Opteron
> 285 instead of the 280?
Yes, that is the question.
>
> Well, its not that you can just go up to a hardware supplier and
> pick exactly the system you want to review/benchmar... especially
> not with pre-production hardware that (at the time) wasn't very
> widely available.
> Normally, you just get what system they have available at their
> marketing or pre-sales department.
Understandable.
>
> The Opteron 280 was from an earlier review and was fitted in the
> "Try and Buy"-version of the Sun Fire x4200. In that system; you
> only have a few options where the 280 was the fastest at the time.

>
> But then again, systems with the Woodcrest 5150 (the subtop one)
> and Opteron 280 (also the subtop one) are about equal in price, so
> its not a bad comparison in a bang-for-bucks point of view. The
> Dempsey was added to show how both the Opteron and the newer
> Woodcrest would compete against that one.

Did I read this correctly that one of the Opterons in the test only
had 4G of ram vs 7 G in the Intel boxes ? If so this is a severely
limiting factor for postgresql at least?

Dave
>
> Best regards,
>
> Arjen
>

On 8-9-2006 15:01 Dave Cramer wrote:
>
>> But then again, systems with the Woodcrest 5150 (the subtop one) and
>> Opteron 280 (also the subtop one) are about equal in price, so its not
>> a bad comparison in a bang-for-bucks point of view. The Dempsey was
>> added to show how both the Opteron and the newer Woodcrest would
>> compete against that one.
>
> Did I read this correctly that one of the Opterons in the test only had
> 4G of ram vs 7 G in the Intel boxes ? If so this is a severely limiting
> factor for postgresql at least?

Actually, its not in this benchmark. Its not a large enough dataset to
put any pressure on IO, not even with just 2GB of memory.

But, to display it more acurately have a look here:
http://tweakers.net/reviews/638/2 and then scroll down to the bottom-graph.
As you can see, the 8GB-version was faster, but not that much to call it
'severely'. Unfortunately, the system just wasn't very stable with that
8GB memory (it was other memory, not just more). So we couldn't finish
much benchmarks with it and decided, partially based on this graph to
just go for the 4GB.

Anyway, you can always compare the results of the Woodcrest with the Sun
Fire x4200-results (called 'Opteron DDR' or 'Opteron 940' in the latest
article) to see how a Opteron with 8GB of memory compares to the Woodcrest.

More of those results can be found in this english article:
http://tweakers.net/reviews/638
And in this Dutch one:
http://tweakers.net/reviews/633

Arjen van der Meijden wrote:
> On 8-9-2006 15:01 Dave Cramer wrote:
>>
>>> But then again, systems with the Woodcrest 5150 (the subtop one) and
>>> Opteron 280 (also the subtop one) are about equal in price, so its
>>> not a bad comparison in a bang-for-bucks point of view. The Dempsey
>>> was added to show how both the Opteron and the newer Woodcrest would
>>> compete against that one.
>>
>> Did I read this correctly that one of the Opterons in the test only
>> had 4G of ram vs 7 G in the Intel boxes ? If so this is a severely
>> limiting factor for postgresql at least?
>
> Actually, its not in this benchmark. Its not a large enough dataset to
> put any pressure on IO, not even with just 2GB of memory.

interesting - so this is a mostly CPU-bound benchmark ?
Out of curiousity have you done any profiling on the databases under
test to see where they are spending their time ?


Stefan

On 8-9-2006 18:18 Stefan Kaltenbrunner wrote:
>
> interesting - so this is a mostly CPU-bound benchmark ?
> Out of curiousity have you done any profiling on the databases under
> test to see where they are spending their time ?

Yeah, it is.

We didn't do any profiling.
We had a Sun-engineer visit us to see why MySQL performed so bad on the
T2000 and he has done some profiling, but that is of course just a small
and specific part of our total set of benchmarks.
Postgresql was mostly left out of that picture since it performed pretty
well (although it may even do better with more tuning and profiling).

We are/were not interested enough in the profiling-part, since we just
run the benchmark to see how fast each system is. Not really to see how
fast each database is or why a database is faster on X or Y.

The latter is of course pretty interesting, but also requires quite a
bit of knowledge of the internals and a bit of time to analyze the
results...

Best regards,

Arjen