IBM T5140 single core performance

Hi all,
I've just done some comparative tests between a Sun Blade 2000
(2x1GHz) and our new T5140 (2x6x1.2GHz), and the single core
performance on the T5140 is somewhat less than I expected. Before I
start dtracing, and all other kinds of tests, I'd like to know if I
should be expecting more from the T5140. Basically, we run Adobe
Distiller Server on both, to distill a large EPS file into a PDF on
the Blade took 8.8 seconds, on the T5140 it took just over 16 seconds.
This would rate clock-for-clock, a core on the T5140 a bit less than
half the USIII in the Blade.

I'll be honest, I expected the single core performance to be not
great, but this is much worse than I thought, particularly as I was
told at a Sun customer briefing to expect clock-for-clock to be much
the same as USIII.

Is this worth investigating further, or does this sound about right?

Cheers

GT

thegman wrote:

> I've just done some comparative tests between a Sun Blade 2000
> (2x1GHz) and our new T5140 (2x6x1.2GHz), and the single core
> performance on the T5140 is somewhat less than I expected. Before I
> start dtracing, and all other kinds of tests, I'd like to know if I
> should be expecting more from the T5140. Basically, we run Adobe
> Distiller Server on both, to distill a large EPS file into a PDF on
> the Blade took 8.8 seconds, on the T5140 it took just over 16 seconds.
> This would rate clock-for-clock, a core on the T5140 a bit less than
> half the USIII in the Blade.

It sounds about right to me. The CPU in T2 processors is much simpler
than an USIII. You could run at least 24 distiller jobs simultaneously
and it would still all be done in 16 seconds, and the slowdown should be
low up to 96 jobs.

> It sounds about right to me. The CPU in T2 processors is much simpler
> than an USIII. You could run at least 24 distiller jobs simultaneously
> and it would still all be done in 16 seconds, and the slowdown should be
> low up to 96 jobs.

I see, certainly we can run many jobs at a time, and get good results
out of the machine, but some are stuck on single cores, and I was told
by a Sun representative that T2+ cores were much better than the T1
predecessor for single core loads, but I'm not noticing any
difference, if anything, it's worse.

I guess if that's the best we can expect, then will need to try to
parallelize the jobs more.

thegman wrote:
> Hi all,
> I've just done some comparative tests between a Sun Blade 2000
> (2x1GHz) and our new T5140 (2x6x1.2GHz), and the single core
> performance on the T5140 is somewhat less than I expected. Before I
> start dtracing, and all other kinds of tests, I'd like to know if I
> should be expecting more from the T5140. Basically, we run Adobe
> Distiller Server on both, to distill a large EPS file into a PDF on
> the Blade took 8.8 seconds, on the T5140 it took just over 16 seconds.
> This would rate clock-for-clock, a core on the T5140 a bit less than
> half the USIII in the Blade.
>
> I'll be honest, I expected the single core performance to be not
> great, but this is much worse than I thought, particularly as I was
> told at a Sun customer briefing to expect clock-for-clock to be much
> the same as USIII.
>
> Is this worth investigating further, or does this sound about right?
>
> Cheers
>
> GT


Does this chip have a decent floating point unit? I gather some of these
chips intended for non-floating point code do poorly if you try to get
them to do it.

Just a thought - I don't know much about this.

thegman wrote:

> I was told by a Sun representative that T2+ cores were much better than
> the T1 predecessor for single core loads,

Say single-thread instead of single-core ;-)

Just for a quick estimate, you can imagine a T2+ core as 2 T1 cores that
got merged, except that they now have a real FPU and an improved crypto
unit. And the clock speed may have been slightly increased. So unless you
are doing crypto or floating points, it isn't that different for a single
thread.

PS: check anything I say with other sources, as I am not an expert.

On Jun 2, 3:48*pm, Marc <marc.gli...@gmail.com> wrote:
> thegman *wrote:
> > I was told by a Sun representative that T2+ cores were much better than
> > the T1 predecessor for single core loads,
>
> Say single-thread instead of single-core ;-)
>
> Just for a quick estimate, you can imagine a T2+ core as 2 T1 cores that
> got merged, except that they now have a real FPU and an improved crypto
> unit. And the clock speed may have been slightly increased. So unless you
> are doing crypto or floating points, it isn't that different for a single
> thread.
>
> PS: check anything I say with other sources, as I am not an expert.

You're right, I should get into the habit of saying "single thread". I
think what we'll probably do is alter our workloads a bit to scale
better, the loads we run which *do* scale get great results on the
T5140, it's just a shame the single-thread numbers are that bad, if
they were more like US-III cores, then the T2+ would be an incredible
processor, right now, I'd say it was just "very good".

thegman wrote:

> I see, certainly we can run many jobs at a time, and get good results
> out of the machine, but some are stuck on single cores, and I was told
> by a Sun representative that T2+ cores were much better than the T1
> predecessor for single core loads, but I'm not noticing any
> difference, if anything, it's worse.

There are various tools which can provide some numbers, like lockstat
and trapstat. They should be a bit easier to use than writing dtrace
scripts.

> I guess if that's the best we can expect, then will need to try to
> parallelize the jobs more.

It all depends on where the jobs get stuck. If there are no hardware
resources, you can't do a thing (assuming you can't rewrite the
application). If it's some kind of software housekeeping, black magick
is usually possible. :-)

--
.-. .-. Yes, I am an agent of Satan, but my duties are largely
(_ \ / _) ceremonial.
|
|