PDA

View Full Version : SB for Benchmarking?



TheMatt
12-20-2002, 12:14 PM
I currently have the opportunity to use/benchmark an Itanium2 box at my work, and I was wondering if sb can be used for this type of benchmarking. As I'm new to SB (I started it up today!), I first don't know if there is a way to use the client as a rough benchmark or not (how long does a usual test take, is there a way to get performance stats out, etc.) . Also, whether or not the Linux client runs under Itanium2 is another question.

Any help from the more knowledgeable people out there? Thanks for anything.

MathGuy
12-20-2002, 12:39 PM
SB is probably not the best choice for benchmarking because the stats are measured in "corrected ExpMod's" (cEM) which depend (mostly) on the particular values for "n" in each block. This measurement is intended to compensate for the fact that the expected length of time taken by a block really does vary. On the other hand, I know that Louie is working on replacing the cEM by some other stat (as yet undetermined) because cEM doesn't do the right thing when "n" varies greatly.

My guess is that the next generation of performance stat will be much more useful for benchmarking than the current one, so you might want to wait for that one.

As to whether the linux client runs on the Itanium 2, my (admittedly speculative) answer would be "not in any way you'd like"...to "go native" the client would *at least* need to be recompiled and the amount of assembly code present would make this task more like a rewrite. I don't know if Louie and kugano have thought about this, but my guess is that there are LOTS of tasks higher on their priority list, given the relatively low numbers of Itanium2 boxen around...

Rumors and rumblings of Itanic x86 emulations are always floating around (mainly because of Hammer) but I really doubt that this would produce the sorts of benchmarks you'd be interested in, anyway.

TheMatt
12-20-2002, 12:47 PM
OK, well, just wanted to ask. As I said, I'm just trying to find some interesting ways to benchmark the new box.

For what it's worth, compiling benchmarks for the Itanium2 has been a nightmare. Frankly, I'll stick with the Alpha for my 64-bit needs until they pry it from my cold dead hands.

rogue
12-20-2002, 01:03 PM
You could try PFGW (http://groups.yahoo.com/group/openpfgw) . It is based on the same code as the SoB client (George Woltman's PRP program) and it has a benchmarking option built in. You will probably be disappointed though with the results because it is optimized for 32 bit platforms, not 64 bit platforms.

Another good choice, which is probably better, would be GMP (http://www.swox.com/gmp). This multi-precision integer package can be compiled on dozens of CPUs and has assembler routines (specific for each CPU) for all the low level stuff. It also has some benchmarking features.

TheMatt
12-20-2002, 01:08 PM
Thanks, I will. One other question that's not that related...what is the best way to gracefully shutdown the client? I mean I can Ctrl-C/kill -15 it easily enough, but do bad things happen when you do that?

funkwad
12-20-2002, 04:21 PM
For what it is worth, I have run both of the linux clients (0.97 and 1.0.2) on an Itanium box with kernel 2.4.9-40SMP.

Emporor
12-20-2002, 06:59 PM
Originally posted by MathGuy
SB is probably not the best choice for benchmarking because the stats are measured in "corrected ExpMod's" (cEM) which depend (mostly) on the particular values for "n" in each block. This measurement is intended to compensate for the fact that the expected length of time taken by a block really does vary. On the other hand, I know that Louie is working on replacing the cEM by some other stat (as yet undetermined) because cEM doesn't do the right thing when "n" varies greatly.


Those stats seemed a little inconsistent. Any idea when cEM will be replaced? I wonder if/how they will convert the current stats. Maybe they will just add an aditional ranking area for the new measurement.

Mystwalker
12-21-2002, 06:27 AM
I guess the new unit will be implemented in the next version of the client. But this could still take some time, as it will basically a more or less complete rewrite AFAIK + some nifty new features.

IMHO, the old cEM value shouldn't be used anymore then, as it led to some confusion when delta n varies greatly between successive work units.
Otherwise, it maybe increases morale when cEMs/sec increases as the project progresses. :D

biff
12-21-2002, 09:09 AM
If you could try to run FreeBench on the Itanium2, it would be very interesting. FreeBench is a cross platform benchmark with results posted for a number of machines. Look at http://www.freebench.org for more information.

If you can, use the Intel v7.0 compiler as I understand it generates very good code for the Itanium2.

/peter

kugano
12-21-2002, 10:51 AM
The replacement unit should be up in the very near future. Clients will still display "cEMs/sec" for rates, but the entire website, including rankings, will change to use the new unit, and the next release of the client (which, yes, is a total re-write and will most certainly have some "nifty" new features).