Disappointing WU times with ver. 3 client...

[Manex]

Fellow Murcers, I am very concerned about our WU output once SETI makes everyone move to the version 3 client. I installed the new client on some HP servers 2 days ago and I have gotten terrible WU times ever since. I have summarized the results below.

(1) angle range: 0.411
cpu time: 6.64 hours

(2) angle range: 0.411
cpu time: 6.715 hours

(3) angle range: 0.417
cpu time: 6.914 hours

(4) angle range: 0.417
cpu time: 6.589 hours

(5) angle range: 0.417
cpu time: 6.911 hours

(6) angle range: 0.491
cpu time: 6.851 hours

(7) angle range: 0.510
cpu time: 7.095 hours

(8) angle range: 0.510
cpu time: 7.083 hours

(9) angle range: 0.509
cpu time: 6.665 hours

I know the new client does additional calculations, and that the number of calculations is dependent on the angle range value. The times above seem consistent with each other. However, what worries me is that these servers process a WU in 3 hours with the version 2.4 client, and the new times are more than double that! I've got some W2K machines at home that were slower than the servers with the version 2.4 client, that are now getting similar times to the HP servers, and that can't be right. If anyone else has seen a similar increase in the WU times, please post here. I hope that there isn't a bug in the HP-UX version of the client... If I don't get a faster WU soon, then I'm going to switch back to the version 2.4 client till they make me switch to the version 3 client.

Eric

[HedsSpaz]

Thats interesting. I for one have been getting phenomenally better times with v3. On my home PC(specs below) I averaged 8 - 8 1/2 hours per wu with 2.4. With v3 I haven't quite pinpointed it down yet but so far I've been averageing between 5 - 6 hours.

Ian

------------------
Epox 7KXA BIOS 5/22, Athlon 650 @ 710, 256 MB PC133 SDRAM, Matrox Mill G400 32 SH PD5.14/PD6.04, SBLive Value,
3Com 3c905B-TX NIC, 56k Zoom PCI winmdm, 13.6 GB 7200RPM UDMA66 IBM HD, 15gb UDMA66 7200RPM Maxtor, 45X Acer CD-ROM, Win 2000, Win 98
RedHat 6.2, MS Intellimouse Optical, 17" Mag 720v^2
Next on upgrade list, Matrox G800? G800 Fusion? DVD drive? new Mobo and CPU? time will tell...


"Any sufficiently advanced technology will be indistinguishable from magic." --Arthur C. Clarke

[Raptor^]

Ok, I'm only guessing here.

Firstly, I think that machines that could fit all of seti v2.4s code and data into their cache will perform worse with v3, as a unit under v3 performs many more operations.
This applies mainly to servers and will be most noticeable if they are relatively slowly clocked.

According to setispy, a WU now needs nearly twice as many floating point operations to process. As I remember, the HP processors were already very efficient at seti 2.4, so do not gain much efficiency from v3, but still have lots more work to do.

A machine that could not fit all of a v2.4 unit in its cache will perform faster - v3 makes more efficient use of the cache than v2.4. This will more than offset the cost of the extra operations, as a single cache miss wastes 100s of CPU cycles. This is most noticeable in high clocked desktops such as mine), they tend not to have enough memory bandwidth for their clock speed and waste even more cycles.

On my box, my fastest unit from seti 2.4 is about 1/5 slower than my slowest from 3.0

Hope that helps (and my ramblings make some kind of sense ). Remember, its just conjecture though.

------------------
P3-700E@933, Abit BF6, G400 MAX, 8.6 gig Seagate, 8.6 gig WD, SBLive 1024, 256Mb PC133... Logitech Optical Wheel Mouse, Keyb, Stuff

[TomNuttall]

I, too, have foung that it takes up to 50% longer to process WUs in V3. as compared to V2.7
P3 650, 512mb ram.

Tom

[agallag]

My P3-650 Thinkpad is considerably faster with 3.0 than it was with 2.4. I'm now averaging about 4-5 hours/WU, when it used to be ~7 hours. My P2-350 and my Xeon 800/133/256kb machines are all pretty much the same as before.

------------------
Andrew

I will not torment the emotionally frail

[Manex]

Up...

Are there any other Unix users out there that have used the version 3 client yet? Please let me know how it performs...

Thanks,

Eric

[Butcher]

On my home machine a WU used to take about 10 hours, it's down to 5-6 now. Seti 3 does make a big difference
As for unix, when (if) seti bring out an AIX version I'll let you know :P

[Kruse]

Hey.

Been running 3.0 CLI at home since it was released. Conclusion: It's definitely slower than 2.4.

It divides like this:

Whenever angle range is <1 WU times are about 1-6 hours (!!!) slower than 2.4.

When angle range is >1 WU times are up to 2 hours faster than 2.4.

The problem: About 9/10 of the WU's I've been getting have angle range <1.

Machine is a PIII-450@558, ram 2-2-2. 2.4 times were around 5.5-6.1 hours. And yes, I've seen several +12 hour WU's since I switched to 3.0.

[Sisyfos]

That's odd.
My observation with CLi 3.0 is:
Angle range <0.85 = same as CLi 2.4.
Angle range >0.85 = faster than CLi 2.4.

Eg.
P3 1050:
<0.85 ~4h30m
>0.85 ~3h30m

P3 300:
<0.85 ~11h30m
>0.85 ~8h30m

Kim

------------------
Join the MURC SETI team!
View the exciting stats

[Kruse]

What OS would that be with Kim?

I now have a substantial result base from both Win98SE and NT4 machines.

The results I mentioned before were on Win98SE, which was the fastest OS for 2.4. It now seems that NT is quite a bit better.

The results I've gotten from my two NT machines have been half an hour faster with 3.0 than with 2.4. Average is based on 100+ WUs.

It also means that my two PIII-500 machines with NT are now faster than my PIII-450@558 with Win98SE. Go figure!

[Maggi]

Hi Guys,

as promised I made some performance observations using V3.0 on my OCed Cel-2 ...

Here's what I got when having my RAM set to CAS3:

angle_range= 0.417
cpu_time=17497.010000 - 4h 51m

angle_range= 0.483
cpu_time=17606.860000 - 4h 53m

angle_range= 0.417
cpu_time=17909.670000 - 4h 58m

angle_range= 0.423
cpu_time=17548.260000 - 4h 52m

angle_range= 0.417
cpu_time=18183.860000 - 5h 03m

angle_range= 0.585
cpu_time=17780.810000 - 4h 56m

angle_range= 0.417
cpu_time=19111.930000 - 5h 18m

angle_range= 0.417
cpu_time=17685.680000 - 4h 54m

angle_range= 0.417
cpu_time=17908.780000 - 4h 58m

angle_range= 0.417
cpu_time=17946.090000 - 4h 59m

angle_range= 0.417
cpu_time=824.600000 - 0h 13m

angle_range= 0.417
cpu_time=18251.750000 - 5h 04m

angle_range= 0.417
cpu_time=17557.760000 - 4h 52m

angle_range= 0.043
cpu_time=24150.350000 - 6h 42m

angle_range= 0.688
cpu_time=16398.010000 - 4h 33m

angle_range= 0.487
cpu_time=17930.710000 - 4h 58m

angle_range= 0.425
cpu_time=18356.870000 - 5h 06m

angle_range= 0.449
cpu_time=18188.030000 - 5h 03m

angle_range= 0.752
cpu_time=17713.150000 - 4h 55m

and here some better (faster) results when having my RAM @ CAS2

angle_range= 7.226
cpu_time=13795.360000 - 3h 50m

angle_range= 0.648
cpu_time=16361.150000 - 4h 33m

angle_range= 0.595
cpu_time=16064.610000 - 4h 27m

angle_range= 3.247
cpu_time=13182.610000 - 3h 40m

angle_range= 0.417
cpu_time=17028.880000 - 4h 44m

angle_range= 0.035
cpu_time=23672.290000 - 6h 35m

angle_range= 0.036
cpu_time=23340.920000 - 6h 29m

... except for those 6.5 hour-WUs at the very bottom, I now can safely say that my WUs hover between 4 and 4.5 hours/WU on average and before V3 it took me around 6 hours/WU.



Btw, those two 6.5hour-WUs have a very small angle range ... is it an incident that those took the longest time to crunch ?

Cheers,
Maggi

[Maggi]

PS: Asus P2B-S, Cel566@952, 192MB PC100, W98 first edition ...

[Brian R.]

Look at it from the overall standpoint. Every group will be hit and/or benefit by this change. I think MURC depends alot on desktop computers and for PIII Coppermines and AMDs, the V3.0 CLI is faster than the previous CLI versions (on average). Some of the teams will be hit very hard owing to their reliance on high-end computers. Not us I think.

In case someone hasn't heard this, the CLI client (V3.0) is much faster than the GUI client for the PIII/AMD computers.

Maggi, I see the same difference with my PIIIs with CAS3 vs CAS2.

[This message has been edited by Brian R. (edited 06 November 2000).]

[CharlesWA]

Look for client version 3.01 soon.

From setiathome.ssl.berkeley.edu

"NOTE: There was a small problem with the Windows 3.0 release and we've temporarily gone back to releasing 2.04. Once compiled and tested, we will release Windows version 3.01."

I have no idea if this will speed things up or not, but at least explains where 3.0 went.

[Sisyfos]

Win2K and NT.