I'm comparing actual operating results of two of my computers against reported BOINC CPU benchmarks.

Computer 1: Win XP 32-bit on 2.8 GHz Pentium-D Dual-Core, 2 threads available (i.e., number of threads = number of cores):
1415 Whetstones per CPU
2403 Dhrystones per CPU

Computer 2: Win 2003 Server 32-bit on two single-core 3.06 GHz Xeon CPUs with multi-threading, 4 threads available (i.e., number of threads is 2x the number of cores):
1442 Whetstones per CPU
3235 Dhrystones per CPU

Now, both are running the same project (Cosmology at home).

Computer 1 takes about 10.5 hours to complete a WU.
Computer 2 takes about 21 hours to complete a WU.

Computer 2 benchmarks slightly faster than computer 1, but computer 2 turns out to be twice as slow in terms of time to complete a WU.

My question is, does the BOINC Manager's "Run CPU Benchmarks" report per physical CPU core, or per THREAD?

It would seem like the former, since computer 2 turns out to be twice as slow per thread in terms of actual runtime required. If not, then something else is causing the discrepancy.

I hope this question makes sense.

Thanks.

Perhaps I provided too much info above.

If you have a machine with twice as many threads as processors, when you run Boinc Manager's CPU benchmarks, and it reports some numbers PER CPU, is it really reporting per A) PHYSICAL processor core, or per B) VIRTUAL thread?

Thanks.

As far as I know, physical core. Benchmarks are only done on CPU zero.

---

I'm comparing actual operating results of two of my computers against reported BOINC CPU benchmarks.

Computer 1: Win XP 32-bit on 2.8 GHz Pentium-D Dual-Core, 2 threads available (i.e., number of threads = number of cores):
1415 Whetstones per CPU
2403 Dhrystones per CPU

Computer 2: Win 2003 Server 32-bit on two single-core 3.06 GHz Xeon CPUs with multi-threading, 4 threads available (i.e., number of threads is 2x the number of cores):
1442 Whetstones per CPU
3235 Dhrystones per CPU

Now, both are running the same project (Cosmology at home).

Computer 1 takes about 10.5 hours to complete a WU.
Computer 2 takes about 21 hours to complete a WU.

Computer 2 benchmarks slightly faster than computer 1, but computer 2 turns out to be twice as slow in terms of time to complete a WU.

My question is, does the BOINC Manager's "Run CPU Benchmarks" report per physical CPU core, or per THREAD?

It would seem like the former, since computer 2 turns out to be twice as slow per thread in terms of actual runtime required. If not, then something else is causing the discrepancy.

I hope this question makes sense.

Thanks.

The cache can make a huge difference in credit thruput, expecially the older xeons that run at 400mhz. The price had really dropped on ebay in the last year and I just bought a pair of SL79V with a 4mb full speed level 3 cache for $48 USD + shipping. Those run at 3ghz and are the fastest 400mhz xeons available. It is significently better then the 2.8ghz "MP" athlon that AMD sold to the server market. I have 13 systems running here and the SL79V is in position 5. Even while running vista 32, it still compares well with the 64 bit dual opterons running linux. Xeons prior to the gallatin did not have a Level 3 cache and intel was losing market share to amd. If your xeon buss is 533mhz then you have more options than me.

I assume your xeon's have a much smaller cache than your dual core. Next time you boot the xeon system, check the intel spec for the cpus and see what size level 2 and level 3 cache they have.

Jord,

Thank you for the answer. This is what I expected, and makes sense given what I'm seeing.

Ed

BeemerBiker,

The machine is an HP DL360 G3 1U-height server that I just HAD to get (it's a piece of Amazon.com history - it was recently retired from the Seattle data center).

That's great info you gave me and it has got me curious as to what other options I might have in the way of processors. Unfortunately, I think the DL360 G3 is very limited to which CPUs will work on its mobo.

Yeah, I know that my CPUs' cache is VERY small (256K) and I see very high soft page fault rates because of it. There is NO L3 cache on them.

Thanks,
Ed

BeemerBiker,

The machine is an HP DL360 G3 1U-height server that I just HAD to get (it's a piece of Amazon.com history - it was recently retired from the Seattle data center).

That's great info you gave me and it has got me curious as to what other options I might have in the way of processors. Unfortunately, I think the DL360 G3 is very limited to which CPUs will work on its mobo.

Yeah, I know that my CPUs' cache is VERY small (256K) and I see very high soft page fault rates because of it. There is NO L3 cache on them.

Thanks,
Ed

That is actually a pretty good system (except for the bottleneck of your cpu). I traded a 15" LCD monitor for a pair of huge Seimens Fujitsu "600" servers that accepted only RIMM memory. I got them for the cases but ended upgradeing the cpu's when I found out about the SL79V chip.

Your system takes SDRAM which is much cheaper than RIMM. There is a list of xeon processors here and your current cpu is probably that Prestonia model RK80532KE083512 but I think you are mistaken about the 256k cache, it must be 512. You might want to double check that as I dont see anything as small as 256 in the 3.06ghz range. you should be able to upgrade to a 533 Gallatin such as the RK80532KE0882M (SL7AE) but they are not as cheap as my 400mhz chip.

Since a 603 chip will fit into a 604 socket, you could put one of the 400mhz Gallatin MP's in your system and run it at 533. I have read where they have been successfully overclocked.

You might wait a while and see if the price drops on the SL7AE and get a pair. It has a 2mb level 3 cache. That will leave a full 1mb L3 of memory for each thread. Currently you have 256kb L2 for each thread. I would only go for the 2mb chip as I dont think that 1mb L3 will provide enough memory for two boinc type tasks. if you are interested in overclocking, send me a private message and we can continue the discussion offline

Beemer,

Your suggestion spurred me into action.

I found a pair of 3.2 GHz 1MB L3 cache Xeons on eBay for $69 and swapped out the old processors. 3.2 GHz is as high as the model G3 will support according to HP's specs. I didn't see any 2MB L3 cache available at the time that had the proper FSB speed.

From what I've seen in the BIOS settings of the G3, it's locked down pretty tight and doesn't allow for much tweaking for OC'ing, and I really don't want to go there as the cooling system is really weak, and the chips run pretty hot already.

The DL360 line has a real, documented problem with fan noise. There are 7 tiny, high-RPM fans that move the air through this slab. The fan speeds are 1) standby (noisy), 2) normal (loud), and 3) high (screaming). With the old CPUs, if I crunch with the Boinc Manager setting over 90%, if the ambient room temp goes over 66F, the fans launch into high and the noise is intolerable.

With my new CPUs, anything above 75% at the same ambient will cause the fans to go into high. I guess the new CPUs run a bit hotter. This is even after several repeat removals and re-installations using different amounts of Artic Silver thermal grease (maybe it was too thin/ maybe it was too thick, etc). Following Arctic Silver's break-in process I've managed to get it up to 75% from 65% before the fans shift into high.

BTW, you were right - my old CPUs were 512K cache, not 256K.

It was an interesting experience, but I didn't plan on the cooling system in this box to be so WEAK. I might end up going back to my old CPUs. We'll see.

Thanks for your suggestions.