After spending some time on SETI Classic to wrap things up, I installed the new V5.2x Versions on Win32 and Linux to have a look what's new.

To my surprise, the Linux boxes *cough* 'still' benchmark almost 50% below that of the Win32 BOINC (?)

Wasn't that supposed to be fixed quite a while ago ?
(I remember the discussions about changing Benchmark routines of the Windows BOINC in order to correct that issue, if my memory serves me right)

I was just ready (with the upcoming new SETI Client) to revert from all optimized BOINC Clients back to the default ones, assuming the heavily assymetric Credit calculations between Linux and Win32 were history, but seems Linux would (again) need those to bring up the numbers at least closer to where they belong...

Having just finished a CPDN Model (which now also indicates Claimed Credits), the Linux V5.24 BOINC System showed a good matching example for the proportion of discrepancy :
Claimed : 3,846.26 | Granted 6,710.74

---

To my surprise, the Linux boxes *cough* 'still' benchmark almost 50% below that of the Win32 BOINC (?)

Wasn't that supposed to be fixed quite a while ago ?
(I remember the discussions about changing Benchmark routines of the Windows BOINC in order to correct that issue, if my memory serves me right)

I was just ready (with the upcoming new SETI Client) to revert from all optimized BOINC Clients back to the default ones, assuming the heavily assymetric Credit calculations between Linux and Win32 were history, but seems Linux would (again) need those to bring up the numbers at least closer to where they belong...

Having just finished a CPDN Model (which now also indicates Claimed Credits), the Linux V5.24 BOINC System showed a good matching example for the proportion of discrepancy :
Claimed : 3,846.26 | Granted 6,710.74

CPDN bases the credit on trickles, so can't really compare. Oh, and a quick look shows that for some of the results it shows Claimed: -- ; Granted 6,805.26 so...

As for benchmark, v4.20 and later should have less difference between platforms, but there's still some disrepancies.


Seti_Enhanced paired with v5.2.6 or later on the other hand doesn't care if the benchmark is off. Example on the claimed credits for the same wu is:
wu-1:
Linux: 60.6420599693264
WinXP: 60.6447411506458
Difference, 0.0044%

wu-2:
WinXP: 60.6513375164549
Linux: 60.6668374576389
Difference, 0.026%


There's very few of the beta-testers that's upgraded to v5.2.6 or later, meaning finding wu's to compare claimed credit between computers is very difficult. There will likely be wu with larger differences in claimed credit, but atleast these two wu indicates much less variation than the BOINC-benchmark is giving...

Hm, so you're saying V5.26 has finally fixed the Problem the V5.22 still has ?

---

Hm, so you're saying V5.26 has finally fixed the Problem the V5.22 still has ?

No, the current SETI@Home-application is still getting the same variation in claimed credit as before.

But, Seti_Enhanced "soon" to be released counts flops, and will therefore have much less variation. You just need v5.2.6 or later to correctly report the flops back to server. ;)

Sounds good, but is that going to be something SETI-Client specific, or will it migrate as a nifty BOINC feature to benefit all other Projects as well ?

Anyway, good to see the old Problem is being finally fixed :)

---

Sounds good, but is that going to be something SETI-Client specific, or will it migrate as a nifty BOINC feature to benefit all other Projects as well ?

Anyway, good to see the old Problem is being finally fixed :)

Any project wanting to use this must do just as with Seti_Enhanced, add code to their application to count flops, or add an application-specific benchmark.


Still, there is one general improvement very resently added to server-code, projects can now set the fraction of floating-point and integer-operations used in their application, and this can give some improvements.