optimized client benchmark results while OC

[mitchellds]

Greetings all,

I noticed that after running the boinc benchmarks after overclocking my amdx2 4400 I actually get worse Dhrystone results. My original results at the default settings were:
2137 Whets
3966 Dhry's

after Oc'ing and bumping up the speed by 200mhz I actually get the below results: notice that the dhry's are down by 1789 but the whetstones are up.

2309 whets
2177 Dhry's

Any ideas why this would be? Are the Dhrystone measurements actually affected by that much by oc'ing, or could the boinc client be reporting incorrect results?

I have a Av8pro with 2 gb of ddr 400 corsair memory, and am running the optimized boinc client posted from this site.

[NeoGen]

That's an odd one...

Have you tested the stability after overclocking? Maybe instabilities can lead to glitches on the benchmark...?

By the way... those values seem kinda low. Are you using the boinc benchmark or some other? Look at my boinc benchmarks from an AthlonXP 2000+.

Measured floating point speed 1619.2 million ops/sec
Measured integer speed 4496.03 million ops/sec

[drezha]

Looks like a BOINC benchmarek but not using an optimised client. ;)

[mitro]

Either force the computer to run the benchmark again (sometimes I get unusually low ones) or the OC isn't stable enough to even handle the benchmark let alone the work. Whatever it is, somethings definately wrong... depending on the client you should be getting much higher benches.

[NVM]

i just checked my Opteron 165, running stock speeds with optimization.


Measured floating point speed 8578 million ops/sec
Measured integer speed 6040.88 million ops/sec

[Strongbow]

I'm no OC'ing expert, Mitro's our guru in that area, but it sounds like the following.

Whetstone is the floating point benchmark so if you speed up your processor then this will be affected.

Dhrystone is the integer benchmark so if you speed up your memory and sysboard (along with processor) then it will increase as it counts the number of loops the test does.

So it sounds like there is a bottleneck probably due to the processor storming along but then hitting contention in the cache due to having to wait for it to read from memory before continuining the next loop. I would suggest checking your memory and sysboard settings to see if they can be aligned with the performance of your stepped up processor.

Could be completely wrong but worth a check!

Mitro needs to step in here to add some experience! :cool:

[AMD-USR_JL]

My drystone goes down a lot too if i oc it too much. Try lowering the base frequency a little, it should help.

[Lagu]

Hi

I have over clocked my computer from 2247/204 MHz to 2310/210 MHz and have this benchmarking result:

2006-04-27 14:37:04|| 3092 double precision MIPS (Whetstone) per CPU
2006-04-27 14:37:04|| 9437 integer MIPS (Dhrystone) per CPU

Lagu :)

[mitro]

Mitro needs to step in here to add some experience! :cool:

Well you asked for it. :D

First of all, people are throwing around 2 different pairs of numbers (or atleast with my client its different). The Whetstone/Dhrystone scores are what are shown if you look in the messages tab of BOINC, while what is listed on the project websites under you computer details is reported as floating point/integer speed (why its different I have no idea).

Now... blackheath, I have to dispute something posted. Memory speed has literally nothing to do with either the Whet or the Drhy. :P

I ran the benchmark twice here at work on my Opteron 175 (OCed to 2640 MHz). The first time I was running 240MHzx11 with the memory running 1:1 and got:

2693 Whetstone
9119 Dhrystone

(This is using the truX calibrating client, by the way)

Then I changed the memory divider to 5:6 (making the memory run at ~200MHz) and got:

2686 Whetstone
9139 Dhrystone

So 40 less MHz on the RAM and the difference was definately not enough to conclude that memory speed (and therefore bandwidth) has any affect on the scores.

Like I said originally, I get those low benches occasionly and I just force it to run again. I use BoincView so I can see if my credits/hour is lower than it usually is.

OT: Opteron 165 coming later today, so I'll have to post tonight or tomorrow just exactly how I OC it. :)

[Strongbow]

Now... blackheath, I have to dispute something posted. Memory speed has literally nothing to do with either the Whet or the Drhy.

You are sort of right Mitro my friend :D , Whetstone is strictly just CPU but Dhrystone is an application that after a few cycles could partially sit in L1, L2 or L3 cache on the processor so it too would not be affected directly by memory performance- unless of course the caches get filled due to lack of coherency and have to wait for validation from main memory. This can often happen in large SMP systems due to bad coding but I'm sure it could happen to the smaller x86/x64 chips as well as they only use MESI protocols rather than MOESI or MOWESI (which my company is about to release in it's next generation of SPARC chips) so they have even less intelligence and would cause the processor to waste cycles during wait for validation. Many things could cause the lack of cache coherency of course and adjusting your sysboard and memory could have affect if there are numerous instructions being performed which are causing the processor to wait several cycles for an answer from main memory.

I must admit though, I'm coming at this from a SPARC background so the above could all be pants but I doubt it!

[Lagu]

Hi

I’ve run Boinc benchmark using normal speed 2247/204 MHz (not over clocked) and the results can confirm Mitro’s conclusion:
There has been little change.

2006-04-27 18:28:50|| 3044 double precision MIPS (Whetstone) per CPU
2006-04-27 18:28:50|| 9361 integer MIPS (Dhrystone) per CPU

When over clock to 2310/210 MHz you can see this:

2006-04-27 14:37:04|| 3092 double precision MIPS (Whetstone) per CPU
2006-04-27 14:37:04|| 9437 integer MIPS (Dhrystone) per CPU

Difference:

MIPS (Whetstone) -48
MIPS (Dhrystone) -76

I think it must be marginally. The question is: How slower is a WU done? 3 to 5 min slower if a WU in the reality takes 12 or 20 hours or only 2 minutes? I think it is depending of which type of WU we crunch. Will there be any faster/slower if we compare uFluids contra QMC@Home or are they equal? Is we could crunch Angstrom atoms or another complex theory model should this affect the speed when thinking of the lesser MIPS?

Lagu :D

[Empty_5oul]

when i overclocked my 64 3200+ by 10% my predictor WU reduced from taking 45mins to just 28mins. My scores went up very little.

[Lagu]

Hi

I’ve checked my team-mates computers benchmarking speed in uFluids and QMC@Home and it is different even among equal processors.

AMD-USR-JL has 2 AMD Athlon 64™ 3800+ Dual core and gets these results:

Measured floating points: 3.3390.31 operations/sec
Measured Integer: 10.664.24 operations/sec

Measured floating points: 3.506.31 operations/sec
Measured Integer: 10.735.94 operations/sec

Mitro has 2 AMD Athlon 64™ 3800+ Dual core and 1 AMD Opteron 4400+ dual core and have these results:

Measured floating points: 3.460.99 operations/sec (3800+)
Measured Integer: 10.337.49 operations/sec

Measured floating points: 3.437.96 operations/sec (3800+)
Measured Integer: 11.229 operations/sec

Measured floating points: 3.025.01 operations/sec (4400+)
Measured Integer: 5.708.48 Operations/sec

Mihost has 1 AMD Athlon 4400+ Dual core 175 and has this result:

Measured floating points 2046.31 operations/sec
Measured Integer: 3.880.83 operations/sec

Why have they with equal processors not the same results? Have they different Boinc versions? Have they different amount of memory? Are they over clocked?

My theory is they got different results from time to time sometimes. And as you can notice, AMD 4400+ have low points and AMD Opteron 175 the lowest. Perhaps Mitro have over clocked one of his computers. I have always been thinking the Opteron should get a much higher score.

Lagu :)

[mitro]

The answer to all these questions is yes. Except when it comes to memory. Memory has ZERO affect on the BOINC benchmark and should have little to no affect on the output of the computer.

And yes my Opteron 175 is overclocked to 2640MHz, but I have a 3800 X2 that runs at 2800MHz 24/7 and my other 3800 X2 is running 2500MHz (but is capable of 2700MHz to make cooling easier since it works as my HTPC).

Blackheath: Now you have me wondering....The benchmark may not be affected by the memory, but the actual crunching might. I've never actually done a lot of testing running the RAM at different speeds and checking the output. Needless to say the difference would be minimal in comparison to the gains in increased CPU speeds.

Well the Opty is here now... time to test the latest dualie!

[Strongbow]

Good point Mitro, apparently some projects like SETI access the main memory a lot. So you may get different performance depending on the type of projects you're testing with!

[Steve Lux]

A lot has to do with how systems are overclocked and how they are optimized. There are many other variables including cooling and how many or what kind of processes that are open.

My system slowed down 10 - 15% with the optimized version for Predictor.