Benchmark and 'sneaker-net' info

[MAD-ness]

I have been doing some 'fiddling' with the client on Windows systems to get some experience with moving around 'tests,' both for the purpose of 'sneaker-netting' and to benchmark the client.

I know that some of you have been able to manually sneaker-net for quite a while now but I finally got around to testing it out. I can move a test to a different computer and get it to run, but the computer is not connected to the net, so it is possible that it would puke if it tried to contact the SoB server. Anyways, I have a small, rough write-up with more specific and perhaps user-friendly directions. I am not posting it publically because I don't really want to be responsible if someone borks thier registry. :)

In regards to benchmarking, I am under the impression that running an identical k,n pair on two machines should allow you to use "cEMs" as a reasonably accurate measurement of performance. The scaling issue with "correction" of the exmods should not apply when dealing with the exact same number, correct?

Here are some initial benchmarks which I ran overnight:


1/23/03 - received from SoB server.

New test exponent for benchmarking:

y2445648

n=2445648
k=33661

Machines:

-------------------
AthlonXP 1800+
ECS K75SA
SiS735 (I think)
512 megs DDR PC2100
Memory timings default
WindowsXP Pro

cEMs/sec: 159599

Run Time: 8:26::10

Full Test: 33.2

-------------------
Athlon Tbird 1.2Ghz
Iwill KK266
KT133A
512 Meg SDRAM PC-133
CAS 2
WindowsXP Pro

cEMs/sec: 98581

Run time: 8:16:38

Full Test: 20.1%

------

I was not trying to prove anything (like architecture advantages, etc.) other than that reasonably accurate benchmarks are possible.

So, I have two questions for y'all:

1) Is my hypothesis that an identical k, n pair should return consistent cEM/s an accurate hypothesis?

2) is it possible to manually create the "zXXXXXX" file and, if so, how?

The next step, IMO, would be to be able to manually plug in a number which the server had not assigned to one of your computers (soley for benching, not to return the result). Editing the registry key (I haven't tried) looks feasible (it shows the format of the values) but that still leaves the "zXXXX" or "yXXXX" (what is the difference?) file on windows as well as the same files on the Linux/FreeBSD systems, plus the "cache" file on the Linux and FreeBSD.

I just tried opening the "zXXXX" file on Windows and it is in binary, apparently. Not gonna be of much use there (unless we can get nebob in here to tick off Louie again). ;)

So, where can I go from here?

[smh]

I think the zxxxxxx files are created by the client.

Just enter the values in the registry and start the client and it creates the file by itself

[MAD-ness]

Thanks.

I had sort of gotten that impression based upon a reference that Louie made one time.

I will test it out when I get home from work.

[key0868]

Could you e-mail me the details on the sneaker-net? I have quite a few machines that I would like to SN files to.

A breakdown of what the function (no PHD level) of each file is would be nice in the meantime also if anyone has a white paper on it.

[key0868]

Disregard.. the above post was intended for DF forum

[MAD-ness]

haha, I just saw a post by you on the Ars forum and had come to that conclusion.

Free-DC also hosts the official DF forum.

http://www.free-dc.org/forum/forumdisplay.php?s=&forumid=27

[Cmarc]

Interesting test MAD-ness.
I run a small Benchmark database for the Ecc2 project. It is a simple tool with two components: a form to input data and a Table to display it.
The value of this kind of ressource is highly dependant on having reproducible results. I thought of moving it over to SoB when I moved the rest of my pages but I didn't believe it would be useful as things stand.
If we had a standard unit of measure to test clients though it could be useful (and current results could be converted should cEMs be replaced at a later date by another unit).
There used to be a 'Benchmark Work Unit' for Seti could we agree on something similar? If so I'd be glad to provide with pages to input and display the data.

Cheers,
Marc

[Mystwalker]

Louie plans to move over to FlOps. :thumbs: But I guess it will take another couple of weeks till we get this nifty feature. :( In the meantime, one *could* do a benchmark with a standard k/n pair, but I'm not sure if it's really that useful then.

So hurry up, Louie, the benchmarkers are waiting! ;)