Check it out OGR counters, 464g client speed improvements!

[Viztech]

I just finished a big WU in my Cel 366@523 running Win 95b

New 464g client actual speed 4.16 Mnodes/sec
Old 463b client actual speed 3.79 Mnodes/sec

That's a 9.7% improvement folks!

The actual vs. benchmark speeds are very close in 2 of my Celerons, much closer than with the 463b client.

I reported earlier a drop in speed in a K6 233 of 5.6% according to the benchmarks. I need to do some actual work units to confirm, but there may indeed be a slight decrease in speed.

viz

 

[Bartman]

where can we get it?

 

[NT4Mike]

Wow, I just benched it and got 7.3 mnodes/sec on my 700@933. Will do a few stubs and see how they come out.

Also, I just pulled a stub from one of Moose's proxies and it was an OGR24 stub

-Mike

 

[Eug]

Yes, much faster in the OGR long bench (Win98), a little under 7 for my Celeron 880, but if I'm interpreting it right it seems to start the stub over from the beginning. If that is the case I shall wait until the end of the current one before I install it.

 

[JonB]

Eug, I restarted with the new client and it didn't appear to discard the old work. It only seems to do that when the full build number changes. These have all been 464 series.

Catch me if you can !!!

 

[Eug]

<< Catch me if you can !!! >>

With your big weekly flushes, you know I can't... for now...

<< Eug, I restarted with the new client and it didn't appear to discard the old work. It only seems to do that when the full build number changes. These have all been 464 series. >>

Ah, I see. With the older 464 betas I was getting a slower speed so I had stuck with the 463b.

 

[Viztech]

Bart-
This is the link to the pre-release page- http://www.distributed.net/download/prerelease.html

Boy, I wish someone would post some Athlon/Duron results.

viz

 

[jinsonxu]

I would try, but i've a phobia of beta clients.

 

[Viztech]

Hehe, I know what you mean about beta software.

What I was doing was copying the beta to either a floppy or a desktop directory. Shut down the regular dnet client, execute the beta from the seperate drive or directory where it cannot goof up the work in progress, configure the beta, and try it out.

I have only let the betas continue to run when I see that I might want to do further testing by completing a work unit or two. The speed improvements on the P2/Celerons have prompted me to install it on several machines at home. It looks like it might be a winner, but would like to see some more actual K6 results as well as the Athlons. This would cover the most popular x86 CPUs.

viz

 

[jinsonxu]

Well, i dunno, i unzipped it. then copied my old ini file over.

But it always gives this message

[Nov 10 05:49:12 UTC] Automatic processor detection found 1 processor.
[Nov 10 05:49:12 UTC] Unable to create network endpoint
ENETDOWN: network down or not available
[Nov 10 05:49:12 UTC] Automatic processor type detection found
an AMD K7-3 processor.
[Nov 10 05:49:12 UTC] RC5: using core #6 (RG/HB re-pair II).
[Nov 10 05:49:12 UTC] RC5: Loaded random 721D2EEC0000000:1*2^28
[Nov 10 05:49:12 UTC] RC5: 0 packets remain in buff-in.rc5
[Nov 10 05:49:12 UTC] RC5: 0 packets are in buff-out.rc5
[Nov 10 05:49:12 UTC] 1 cruncher has been started.

But my old client doesn't have any problems accessing my proxy.

 

[jinsonxu]

Turns out that it has troubles accessing my proxy. Main keyservers are ok. Any ideas?

Here's the long bench for the new client.
[Nov 10 05:58:39 UTC] #1: OGR:24/12-28-4-2-3 [130,335,027]
[Nov 10 05:59:00 UTC] OGR: Benchmark for core #0 (GARSP 5.13)
0.00:00:17.82 [7,001,461 nodes/sec]
[Nov 10 05:59:20 UTC] #1: OGR:24/12-28-4-2-3 [258,594,387]

The old client's bench

[Nov 10 06:00:39 UTC] 1 cruncher has been started.
.....10%.....20%.....30%.....40%.....50%.....60%.....70%R
[Nov 10 06:01:02 UTC] Benchmark for OGR core #0 (GARSP 5.13)
0.00:00:16.03 [7,160,709.94 nodes/sec]
.....10%.....20%.....30%.....40%.....50%.....60%.....70%R


Both benches done under the same conditions, one after the other. Loads of windows opened, proxy running, webserver running, and ICQ messages queued.

 

[jinsonxu]

Here's the RC5 benchmark for the new client

Long
0.00:00:20.07 [3,203,864 keys/sec]
Short
0.00:00:08.04 [3,205,158 keys/sec]
Uses this core (RG/HB re-pair II)

For the old client,

Long
0.00:00:16.12 [3,186,197.19 keys/sec]
Short
0.00:00:08.23 [3,182,542.18 keys/sec]
Uses this core (RG/HB ath).

 

[Viztech]

Hmmm, that is not good.

I had a little trouble with the first 464 beta connecting to my LAN proxy, but that was quickly resolved in the next versions.

A couple of questions.
Are you trying to reach your proxy on your LAN?
Are you using the proxy machine's IP address?

viz

 

[jinsonxu]

So on my Duron,

The new client is slower in OGR benches by 2.22391 %

But it is faster in RC5 than the old one by 0.710626 %

 

[jinsonxu]

viztech, i solved that problem. Turns out that the new client doesn't seem to accept jinsonxu.teamanandtech.com.

I changed it to the IP address and it's working fine now!

I'm now waiting for my 2nd stub to complete. THere's a weird result from my first one. 0.00:03:32.36 - [5.78 Mnodes/s]???

 

[Viztech]

Well, this could be a problem, as it seems that your client is not getting DNS resolution to go onto the Internet, get an IP address to come back to your proxy.

Is this client behind a firewall or using Network Address Translation, like Windows ICS?

edit I was able to get to your proxy getting thru my firewall, so we know that DNS is working for teamanandtech.com
[Nov 10 06:29:41 UTC] The perproxy says: &quot;HuanZhuFans@TA Pproxy!!&quot;
[Nov 10 06:29:41 UTC] Retrieved project state data from server. (cached)
[Nov 10 06:29:41 UTC] OGR: Retrieved packet 1 of 1 (100.00%)
[Nov 10 06:29:41 UTC] Connection closed.
[Nov 10 06:31:06 UTC] #1: OGR:25/6-4-16-2-7-27 [40,723,880,272]
/edit

 

[jinsonxu]

Firewall. BlackIce.

Does the speed that OGR process get bogged down so much by normal usage of the computer? It's far from the benchmark.
There's something weird. Think i should restart my computer?


I'm waiting for the next stub to be completed. The first one finished rather faster than i expected. Which stubs are the longest to complete?
I'm dying to see OGR stats. PPstats i mean.

EDIT:It's been 2 stats units and the rate is still 5 plus only!
0.00:16:15.28 - [5.63 Mnodes/s]
[Nov 10 07:07:43 UTC] OGR: 12 packets remain in buff-in.ogr
[Nov 10 07:07:43 UTC] OGR: 2 packets (8.79 stats units) are in buff-out.ogr
[Nov 10 07:13:01 UTC] #1: OGR:24/12-28-4-6-19 [1,898,952,987]

 

[Viztech]

Might be a conflict with Black Ice and the new network layer on the client.

This could be a real problem with Round Robin users. Could you please post a description of your setup? We surely have some Black Ice users who can help out on this problem.

Well, I guess that it's good to see these problems now while it's in Beta!

As for the benchmarks, I saw the older 463b client bench ~1% slower than actual output. That helps take up some of the difference of that 2.2% loss that you saw.
BTW, were you running the 463b client with your old benches?

viz

 

[jinsonxu]

Well, i disabled BlackIce and there was no problem. Since you say that you cna get through even with my firewall up, i wouldn't worry.

Yes, the old client i'm using is 463b

 

[ViRGE]

There isn't a Black Ice problem as far as I can tell. I have Black Ice and the latest client, and its working fine.

 

[jinsonxu]

It isn't. It's just that my black ice settings are paranoid.

 

[Eug]

Celemine 880

OGR: Benchmark for core #0 (GARSP 5.13) [6,947,468 nodes/sec]
OGR: Benchmark for core #0 (GARSP 5.13) [6,945,916 nodes/sec]
OGR: Benchmark for core #0 (GARSP 5.13) [6,943,976 nodes/sec]
OGR: Benchmark for core #0 (GARSP 5.13) [6,945,528 nodes/sec]
OGR: Benchmark for core #0 (GARSP 5.13) [6,946,304 nodes/sec]

Average 6945838.4 nodes/sec = 7893 nodes/sec/MHz.


Celemine 897

OGR: Benchmark for core #0 (GARSP 5.13) [7,076,784 nodes/sec]
OGR: Benchmark for core #0 (GARSP 5.13) [7,074,771 nodes/sec]
OGR: Benchmark for core #0 (GARSP 5.13) [7,076,784 nodes/sec]
OGR: Benchmark for core #0 (GARSP 5.13) [7,078,396 nodes/sec]
OGR: Benchmark for core #0 (GARSP 5.13) [7,075,173 nodes/sec]

Average at 897 is 7076381.6 equals 7889 nodes/sec/MHz.

My previous at this speed is 6968 nodes/sec/MHz.


WooMoo 7 Mnodes/sec!!! This is a 13% speed increase for me over the 463b client people.

(P.S. Works fine with BlackICE Nervous setting.)

 

[Eug]

I guess we should point out now that PIII/Celeron rivals the Athlon Thunderbird core for OGR speed, on a MHz per MHz basis. As I've illustrated, the Celeron 900 does nearly 7.1 Mnodes/sec. jinsonxu is getting almost 7.2, but with a Tbird at 1 GHz. (He probably could get a bit more if he shut off all his open apps though.)

So, the calculations we did way back about appropriateness of various CPUs for which project are now somewhat out of whack.

 

[jinsonxu]

Well, no.

I'm ran my Duron at 900Mhz when performing this test. Still am cause this chip only requires 1.55V at this speed and i decided that running 900Mhz at lower voltage 24/7 was more desirable.

 

[Eug]

Wow, that was quick!


<< I'm ran my Duron at 900Mhz when performing this test. Still am cause this chip only requires 1.55V at this speed and i decided that running 900Mhz at lower voltage 24/7 was more desirable. >>

Heheh, I'm running 1.9 V 24/7 simply because I like knowing I can do 7 Mnodes/s. (I only need 1.8 V to run the 880 MHz, and probably would be OK running 1.85 V at 900+, but my motherboard doesn't support that voltage. ).

Anyways, I dunno why I read Tbird. My bad. However, it still seems like a good comparison of the Celemine vs Duron. Based on this I'd guess a Duron is on average 2% faster. With the old client, a Duron would leave a Celemine in the dust.