speed predictions...

[Umfriend]

OK VJ, congrats on passign me for the very last time. Enjoy it while it lasts.....

One thing on yer setup: *if* HT was on just as long as it was off, then my guess is that the s/TS going from 7.5 to 5.4 actually means you've been running at:
5.4*2 - 7.5 = 10.8 - 7.5 = 3.3 s/Ts since HT was turned off.

This is still slow. So yeah, the info requested above would be usefull as I *think* you ran with HT turned on quite a bit longer than with it turned off.

Meanwhile, I am concerned about James_D being gone for over 2 weeks now and Rattledagger either having 900 or having 30,000 credits. What's up with that dude?

[Umfriend]

Well, doing the numbers, you're crunching at apporx 3.55 s/Ts, on each CPU. I serioulsy think you are mem-bandwidth limited.

[dbdg]

I'd say so too, my PIV 3Ghz goes at 1.9 Sec/TS, 3.2 for my 1.8.

Annoyingly I realised that my main rig was paused for nearly 20 hours yesterday

[Umfriend]

VJ, I'm done

[VJ]

Ok, this is the mainboard:

http://www.supermicro.nl/products/motherboard/Xeon/E7505/X5DA8.cfm

I have 1 GB (2x512 MB) of DDR266 registered ECC memory. The Xeons have 512 KB of Level2 Cache (each).
In reviews, I remember reading that the memory that is registered tends to be slower (1 additional step needed for every memoryaccess) than unregistered or buffered memory. However, as the mainboard only supports registered, I had no choice.


The first time, HT was enabled, and I ran 4 units at the same time. I don't know how long it ran anymore... but I suspect it couldn't have differed more than 20 hours.
After the first week (+ perhaps 2 days, hence the 20 hours), I uploaded and got just over 3000 credits. The indicated speed was 7.5 sec/TS. The prediction for each unit was 1600 hours.

Now, I had HT still enabled, but BOINC was limited to only using 2 CPUs. After upload yesterday, I got just over 2000 credits, with a speed of 5 sec/TS. The predictions for the running units was the same, but for the paused units is was about 650 hours.

Both the changed prediction as the speed seem to indicate I crunch faster...


Jörg

[Umfriend]

I know nothing about HT, but doing it like that, don't you risk running the two units at one HT'ed CPU?

Not that it neccessarily makes a lot of difference. I'm checking now whether the mem is dual channel, hmm, it is apparantly. It's too bad yer computers are not visible, but the math is really not that difficult. You need to work off of the actual recently uploaded results though as the 7.5 and 5.4 are overall averages (see above) and thus are useless.

How many Computer IDs did you get for this one, 1, 2, 4? If the latter, go to your account, your computers, and the select Last Trickle for each. You'll get a list with more than 10 trickles, copy those neatly into a spreadsheet and mail to my Umfriend_1 mail account at yahoo dot com. I'd love to see if I can make some sense out of that.

I still believe you are bandwidth limited. My DIL's PC (which is offline for a few days, but crunching, I can pull that stunt as well) is a P4-2.66 (not 2.4 as I first thought). It's the 505 (no HT, 1 Mb L2 Cache), with DDR333 dual channel Ram. It's crunching at 2.1 s/Ts. Now, speedwise, it ain't that much faster, but per CPU, the mem bandwidth is way over twice as hign. Think that is the culprit.

You *might* want to change it to running at 1 CPU only for a day and see how many CPU secs it took for one timestep. Would not be surprised if you got it down to, say, 2.4 s/Ts.

Overall, you'd still be slower than doing it duallie, but 2 cheap HP's would own ya

[VJ]

Originally posted by Umfriend
I know nothing about HT, but doing it like that, don't you risk running the two units at one HT'ed CPU?

I'm running XP, which is hyperthreading aware. Consequently, it will schedule the process in such a way that performance is optimal. (I can verify that they run on 2 physically different CPUs)

Not that it neccessarily makes a lot of difference. I'm checking now whether the mem is dual channel, hmm, it is apparantly. It's too bad yer computers are not visible, but the math is really not that difficult. You need to work off of the actual recently uploaded results though as the 7.5 and 5.4 are overall averages (see above) and thus are useless.

The computer can never be visible, as I use an other one to upload the data. I can only see the last couple of trickles, so I can only give these last results.

How many Computer IDs did you get for this one, 1, 2, 4? If the latter, go to your account, your computers, and the select Last Trickle for each. You'll get a list with more than 10 trickles, copy those neatly into a spreadsheet and mail to my Umfriend_1 mail account at yahoo dot com. I'd love to see if I can make some sense out of that.

It was recognized as the same PC. It just changes IDs if I change the preferences (i.e. the number of CPUs that can be used by BOINC).

You *might* want to change it to running at 1 CPU only for a day and see how many CPU secs it took for one timestep. Would not be surprised if you got it down to, say, 2.4 s/Ts.

Problem is that uploading is quite a hassle... I might try that, but now is too busy a period to do so.

I wonder if the read-outs are affected by using another computer to upload/download...

This is what I can find about my last upload:

Time Sent (UTC)	Host ID	Result ID	Result Name	Phase	Timestep	CPU Time (sec)	Average (sec/TS)

28 Nov 2004 13:37:07	hidden	380790	3jfd_100186805_0	1	216040	1177432	5.4501	
28 Nov 2004 13:37:07	hidden	380790	3jfd_100186805_0	1	205238	1139317	5.5512	
28 Nov 2004 13:37:07	hidden	380790	3jfd_100186805_0	1	194436	1100732	5.6612	
28 Nov 2004 13:37:07	hidden	380790	3jfd_100186805_0	1	183634	1062365	5.7852	
28 Nov 2004 13:37:06	hidden	380790	3jfd_100186805_0	1	172832	1024118	5.9255	
28 Nov 2004 13:37:06	hidden	380790	3jfd_100186805_0	1	162030	985895	6.0846	
28 Nov 2004 13:37:06	hidden	380790	3jfd_100186805_0	1	151228	944915	6.2483	
28 Nov 2004 13:37:05	hidden	380790	3jfd_100186805_0	1	140426	903664	6.4352	
28 Nov 2004 13:37:05	hidden	380790	3jfd_100186805_0	1	129624	863558	6.6620	
28 Nov 2004 13:37:05	hidden	380790	3jfd_100186805_0	1	118822	823908	6.9340

(for some reason, it puts a newline on a weird place. That last digit of each line is the first digit of the CPUtime)
It does appear to be dropping...



Jörg

[Umfriend]

Yeah, I used those numbers to come up with a 3.55 s/Ts but then two units, so effectively 1.775 s/Ts in single CPU or unit terms. This is fast, but far slower than two cheap HP's (800 euro a piece, no display). I think cost/performance wise for CPDN, duallies are not the way to go unless you can supply them with a huge mem bandwidth.

Sry dude, you're being owned by the Umf!

[VJ]

Originally posted by Umfriend
Yeah, I used those numbers to come up with a 3.55 s/Ts but then two units, so effectively 1.775 s/Ts in single CPU or unit terms. This is fast, but far slower than two cheap HP's (800 euro a piece, no display). I think cost/performance wise for CPDN, duallies are not the way to go unless you can supply them with a huge mem bandwidth.

My P4 2.4 averages 2.54 s/TS, so the 1.775 s/TS for a 2.4 GHz Xeon is not bad. Which HP are you comparing to?

But, I can still run heavy software on my PC, while it is working on 2 units in the background...

Too bad that my upgrade scheme was foiled by Intel when the released the 800 MHz FSB Xeons.

Sry dude, you're being owned by the Umf!

by Umf's family you mean?
Statistically, I crunch more or less the same with only one quarter of the CPU power. And DW can come into play any minute...


Jörg

[Umfriend]

You can't seriously compare the 1.775 s/Ts with the 2.54 of the single CPU machine. If you have to, then it compares badly as two P4 2.4 machines would yield you 1.27 s/Ts in same terms....

The client runs on low priority, so yes, it does not interfere with normal work even if you have set preferences to run when the computer is in use even, just as long as you don't run out of physical memory I guess.

I'm comparing to a HP Pavilion a760 btw.

No, you crunch far less. It took me one week to catch up with you, so that was about 3,000 credits more than you did, including your weekly upload

I told my sister to start turning the machines off soon BTW, I think this is costing here about 1.71 euro's a day ( assume 100W of consumption on here PC's).

[VJ]

Originally posted by Umfriend
I'm comparing to a HP Pavilion a760 btw.

I meant: what CPU is in that...

No, you crunch far less. It took me one week to catch up with you, so that was about 3,000 credits more than you did, including your weekly upload

I know...
At least I had the joy of being in the top 10, even if it was only for a small part of the day.
BTW, do we get bonus credits when a workunit finishes? I so, expect my credits to rise within the hour...

You know: with 8 PCs, you should be in the top 2... Go on, try to pass Maggi!
Always picking on the little guys, what chance to I have when it is 8 against 2? Pick on someone your own size!

I told my sister to start turning the machines off soon BTW, I think this is costing here about 1.71 euro's a day ( assume 100W of consumption on here PC's).

Add it to her Christmas present?
And what about the team?
At the current rate it will be hard to climb further, but if people start switching off computers just because they are ahead of me! You, sir, should be ashamed of yourself!

j/k

Currently, there are electricity problems (actually, meter problems) in my appartement, so for the time it doesn't matter that much. My PC has this PSU:

http://www.pcpowercooling.com/products/power_supplies/maxperformance/turbocools/index_hp_atx-3.htm

(510 XE: 510 W, 650 W peak )
But I'd be suprised if it used that much while crunching (few diskaccesses, ...).

On the positive side: I could actually turn my heating down quite a bit...



Jörg

[Umfriend]

I meant: what CPU is in that...

Well, I did say

My DIL's PC (which is offline for a few days, but crunching, I can pull that stunt as well) is a P4-2.66 (not 2.4 as I first thought). It's the 505 (no HT, 1 Mb L2 Cache), with DDR333 dual channel Ram. It's crunching at 2.1 s/Ts.

What else do you need, serial no.?

Yeah, I'll start picking on anyone I think I can pass once I have a good view on how much credits I'll (or we'll) be running going forward. Don't want to start a loosing battle, and against you, it took all, really, all I could manage.

You're right, but maybe it would help if we got more members as well. Try and get dZeus and Technoid to join, will ya?

Could be btw, that my sis will 9i]refuse[/i] to switch them off, LOL.

[VJ]

Originally posted by Umfriend
Well, I did say
What else do you need, serial no.?

And where exactly did you post the link between DIL and a HP Pavilion a760 ?

Yeah, I'll start picking on anyone I think I can pass once I have a good view on how much credits I'll (or we'll) be running going forward. Don't want to start a loosing battle, and against you, it took all, really, all I could manage.

You mean you'd stop nagging after you passed me? Aarghh, why did I put up a fight?

Euhm, are the workunits counting down?
My previous workunit started with 3ev9, my current one starts with 22el...


Jörg

[Umfriend]

And where exactly did you post the link between DIL and a HP Pavilion a760 ?

ooops, sry

[Umfriend]

Well VJ, rejoice!

Sis is saying the room where her 4 PC's are was getting warm She turned them off, and now onl run when actually used. This w/end you'll have passed me I'd think.