Cacarulo: Interesting and upseting experiment

[Cacarulo]

This post is for those of you who have an Intel PIV Prescott 3.0 Ghz with HT and wanted to try the following experiment:

Yesterday I needed to run a sim with SBA (10,000,000,000 rounds) single deck, and found that the estimated time for achieving this job was 44 hours. I thought this was too much so I decided to run the same sim on a PIV 2.8 Ghz (no Prescott and without HT). To my surprise the estimated time went down to approx. 22 hours! This means in other words that for sims my PIV 2.8 is twice as fast than my PIV 3.0 HT.
Hey, what's going on here? So, I do another experiment: I disabled the HT option from the motherboard setup, run the sim again and voil?: 30 hours!
Still worse than my PIV 2.8 Ghz but better than with HT enable.
Can anybody explain the reason of this behaviour?

Sincerely,
Cac

[Magician]

> This post is for those of you who have an
> Intel PIV Prescott 3.0 Ghz with HT and
> wanted to try the following experiment:

> Yesterday I needed to run a sim with SBA
> (10,000,000,000 rounds) single deck, and
> found that the estimated time for achieving
> this job was 44 hours. I thought this was
> too much so I decided to run the same sim on
> a PIV 2.8 Ghz (no Prescott and without HT).
> To my surprise the estimated time went down
> to approx. 22 hours! This means in other
> words that for sims my PIV 2.8 is twice as
> fast than my PIV 3.0 HT.
> Hey, what's going on here? So, I do another
> experiment: I disabled the HT option from
> the motherboard setup, run the sim again and
> voil?: 30 hours!
> Still worse than my PIV 2.8 Ghz but better
> than with HT enable.
> Can anybody explain the reason of this
> behaviour?

Estimated time? Could it just be that the estimates are inaccurate on machines with hyper-threading? Or did you actually run the sim to completion on both machines? Is the amount and type of RAM in the machines identical? Is there other software running on the machines?

Hopefully some other lucky individual with a P4 3.0GHz will offer to run the sim for you to compare results.

[Cacarulo]

> Estimated time? Could it just be that the
> estimates are inaccurate on machines with
> hyper-threading?

Don't know.

> Or did you actually run the
> sim to completion on both machines?

No needed to. It was a lot of time to wait for.

> Is the
> amount and type of RAM in the machines
> identical? Is there other software running
> on the machines?

Similar software but different RAM. PIV 3.0 is 512Mb/400Mhz and PIV 2.8 is 256Mb/333Mhz.

> Hopefully some other lucky individual with a
> P4 3.0GHz will offer to run the sim for you
> to compare results.

And how can you explain HT vs no HT differences?

Sincerely,
Cac

[Norm Wattenberger]

Odd. CVData runs about the same speed with or without Prescott with or without HT. It must have something to do with the fact it's a DOS program. Hit ctrl-alt-del and look at the performance tab. If it is running only on one CPU, that would seriously hurt performance.

[Cacarulo]

> Odd. CVData runs about the same speed with
> or without Prescott with or without HT. It
> must have something to do with the fact it's
> a DOS program. Hit ctrl-alt-del and look at
> the performance tab. If it is running only
> on one CPU, that would seriously hurt
> performance.

But if I disable HT it would be the same as running on one CPU of 3.0 Ghz, no? My other Pentium (2.8) is still faster.

Cac

[Norm Wattenberger]

> But if I disable HT it would be the same as
> running on one CPU of 3.0 Ghz, no? My other
> Pentium (2.8) is still faster.

I was looking at HT vs. non-HT. As for Prescott vs. non-Prescott, only things I can imagine are the fact it's a DOS program, or the way it uses cache or arithmetic pipielines. But I'm just guessing.

[Magician]

> No needed to. It was a lot of time to wait
> for.

Maybe it will actually finish much faster than the estimated time says. Perhaps the code SBA uses to calculate the remaining time is not accurate with HT enabled?

I'm probably wrong but it might be worth running 200,000 rounds or so to make sure.

> And how can you explain HT vs no HT
> differences?

I think it's been said that hyperthreading can actually reduce performance in some applications.

All other things being equal, the 3.0GHz CPU should definitely beat the 2.8GHz one. So some other things most not be equal.

[Magician]

> Hopefully some other lucky individual with a
> P4 3.0GHz will offer to run the sim for you
> to compare results.

Well, my brand new P4 540 (3.2GHz with HT) arrives this weekend. I could run the sim for you some time next week if you're still interested. Just drop me an email.

[Cacarulo]

> Well, my brand new P4 540 (3.2GHz with HT)
> arrives this weekend. I could run the sim
> for you some time next week if you're still
> interested. Just drop me an email.

Yes please, I would like to see your results.

Cac

[mbman]

It is almost certainly an error on the part of the s/w, as the s/w cannot be aware nor can it calculate the efficiencies of the HT CPU. That being said, it might be more useful to run the program 4 or more times at the same time on the HT machine and combine the results in the end, as that is what the HT advantage is supposed to be, the ability to run multiple applications ('threads') at the same time faster.

> Maybe it will actually finish much faster
> than the estimated time says. Perhaps the
> code SBA uses to calculate the remaining
> time is not accurate with HT enabled?

> I'm probably wrong but it might be worth
> running 200,000 rounds or so to make sure.

> I think it's been said that hyperthreading
> can actually reduce performance in some
> applications.

> All other things being equal, the 3.0GHz CPU
> should definitely beat the 2.8GHz one. So
> some other things most not be equal.

[Phinitum]

HT allows 2 threads to run on the same core. In the best case there is a small improvement in throughput, maybe 10%. What this indicates is that the two threads are almost always blocking each other. And several things can go wrong, particularly when the OS or application isn't aware that these aren't real processors.

With one processor known, a lower priority task will get scheduled when the higher priority task is in a wait state. With HT the high priority task can get slowed down by almost 50% when another task gets dispatched.

The OS may allow two competing tasks on the HT processors to keep invalidating each others cache, resulting in much more memory access than without HT.

An application that is multi-processor aware may break up its work into 'jobs' for each processor, this may result in more total processing than would be needed for a single thread, but require less elapsed time if the processors each go full speed. MS SQLServer has this problem sometimes, generally an application has to be pretty sophisticated for this to be the case.

> It is almost certainly an error on the part
> of the s/w, as the s/w cannot be aware nor
> can it calculate the efficiencies of the HT
> CPU.

[Magician]

> Yes please, I would like to see your
> results.

OK, I'm ready to go. Is there a specific sim you want me to run?

[Cacarulo]

> OK, I'm ready to go. Is there a specific sim
> you want me to run?

Could you try to reproduce what I posted in my first post of this thread? Don't need to be exactly the same details of my sim. If you don't have two different processors just try the HT vs No HT test.

Thanks in advance.

Sincerely,
Cac