OGR: Linux vs. Win9x speeds

[sciencewhiz]

Several people have asked for benchmarks between version 462 and 463 of the dnet client and between Linux and Win9x.

My setup is a PIII 500 with 128mb of ram that tri-boots Win98SE, Caldera Openlinux and BeOS PE. I downloaded and installed version 462 and 463 of the distributed net client for both Win9x and Linux. The windows benchmarks were run with only explorer and systray running. In linux, I ran the benchmarks from the command line. I ran 6 long benchmarks for each client version (462 and 463) and each OS (Win98SE and Linux). I then averaged the results and also found the standard deviation in the mean for each platform and version. This way (even though it was only six trials) I could state the results as the average +/- the standard deviation in the mean.

The results:

Win98SE, RC5 with version 463b: 1,405,400 +/- 600 keys/sec
Win98SE, RC5 with version 462: 1,405,800 +/- 600 keys/sec
Linux, RC5 with version 463: 1,405,400 +/- 300 keys/sec
Linux, RC5 with version 462: 1,405,400 +/- 400 keys/sec

As you can see, there is absolutly no difference between rates.

Now for the OGR rates:

Win98SE, OGR with version 463b: 3,374,000 +/- 4000 nodes/sec
Win98SE, OGR with version 462: 2,737,000 +/- 3000 nodes/sec
Linux, OGR with version 463: 2,578,000 +/- 1000 nodes/sec
Linux, OGR with version 462: 3,183,600 +/- 400 nodes/sec

As you can see, there are wide variations between the rates based on OS and client version. In linux version 462 is 23% faster than 463 but is still 6% slower than 463b in Windows.

The other thing of interest is that in linux the standard deviation of the rates is much lower than in Win9x.

If you would like me to do these benchmarks for more versions or for BeOS let me know and I'll run them when I have time.

Edit: I used 463b in windows.

 

[RC]

whiz: Thanks for verifying this. From OGR/Linux/462 client to OGR/WIN32/463b client my numbers showed about 7% and yours showed a 6% difference. The one percent difference is not important. So I calculate I'm loosing approx. 350 Gnodes/day or about 1Tnode/3 days by using Linux/462 client.

My crackrack nodes each contain 64MB. I think I can run WIN 98se with 64MB, but I'm not certain. However, all the nodes in my crack rack are diskless. I use a Linux distro and boot from a floppy.
To regain the 350Gnodes that are lost each day I would need to add 6 HDDs. From looking at pricewatch the cost would be around $200 for 6 HDD. If I go the used HDD route I risk the potential for additional trouble.

The release date for the Linux 463 client (July 13, 2000) is the same as for the 463b version of WIN32. So why is the OGR/Linux/463 client 30% slower than the OGR/WIN 32/463b client? How can this be?

What really has me frustrated is that I have reported this oddity to dnet and they have chosen to ignore my email.

 

[sciencewhiz]

Damn, you have a lot of power. You're losing about the same amount as I am cracking.

64 megs is more than enough to run win98se. But, you would have to get some hard drives.

Have you considered using the dos client? You could boot them using a dos boot disk and then have another disk with the client and a batch file that sets up a ram drive and starts the client cracking. The only disadvantage (that I can think of) is that you would have to sneakernet the clients.

 

[RC]

DOS - no LAN support, poor performance, shall I go on

I am rather upset with the lousy alternatives.

How long before dnet gets off their arse and updates the Linux client? I suppose I better not wait on that happening.

So it seems that my original thoughts are still correct. Either live with it, or purchase 6 HDDs.

🙁🙁🙁

 

[RC]

Synopsis: Diskless Linux nodes and OGR are not a great combination.

Everyone agree?

 

[Viztech]

RC-

The other things that concern me with the diskless clients is that you would loose your work in the event of a power failure. :-(
If you have check points enabled with a hard drive, your PARTIALLY completed packet will be saved.
Ummm, I might propose that you use Hard drives, Win 95b (aka OSR2) and the 463b client.

YGPM

viz

 

[Joe O]

I've seen this result in other distributed client efforts as well. The consensus is that the Windows compilers do a better job of optimization than the Linux clients. You have to go to hand optimization and/or assembler, to do a better job.

 

[RC]

Joe your explanation is the only answer that explains why the release dates for Linux 463 and WIN32 463b are the same yet there is a 30% difference in performance.

Your explanation does not explain why the previous version of the Linux client (462) is 23% faster than the latest Linux client.

 

[Joe O]

RC, Actually it does. There are a number of possibilities. Two that immediately come to mind are:
1)A small change in the source code prevented the compiler from optimizing the code as well. It doesn't take much. I've seen comments do it! This code change could even be the one that resulted in the Windows improvement.
2)They hand optimized the previous client, but did not optimize this one.

 

[RC]

1. "small change in the source code" - I did not consider the fact that they may have changed the code without any testing to verify results. Are you saying that they operate in such a manner? Incredible.

2. "They hand optimized the previous client, but did not optimize this one"
Since the popularity of Linux has increased I would hope and think that the oppposite would be true.

Let's look at the facts:
1. Dnet only has two projects RC5 and OGR
2. Linux is currently one of the most popular operating systems
So why would they release a client for Linux that reduces OGR performance by 23%.

They are understaffed. This is a given. But this severity of a screw up is mind boggling.

I'm not flaming you, but the fact that they are so inept, do not return emails,... has my blood boiling.