| Index | Recent Threads | Unanswered Threads | Who's Active | Guidelines | Search |
 |
World Community Grid Forums
Category: Support
Forum: BOINC Agent Support
Thread: Much smaller performance under linux? |
| No member browsing this thread |
|
Thread Status: Active Total posts in this thread: 18
|
|
| Author |
|
|
Former Member
Cruncher Joined: May 22, 2018 Post Count: 0 Status: Offline |
I used Slax 6.0.7, so its a live linux.
I didn't use a command to see the running processes, instead I took the "Performance Monitor (KSysGuard)" to see the current CPU-usage and clock. There was almost no CPU-Usage when boinc did nothing. When BOINC did its benchmarks, CPU usage increased to 100%. |
||
|
|
Former Member
Cruncher Joined: May 22, 2018 Post Count: 0 Status: Offline |
The benchmarks are useless, ignore them. This has been a problem since the beginning of the BOINC Project and was actually much worse prior to BOINC 5.8. All the individual projects know this and compensate for the skewed benchmark results in some way. Basically it is caused by the compiler used under Windows over-optimising the benchmark code in the BOINC client, while the one used under Linux and all other platforms doesn't. So actually the Linux results are accurate but the Windows ones are inflated.
----------------------------------------The simplest method of fixing this would be to use the same compiler (and compiler options) for the BOINC client on all platforms, or move the benchmarking code inside the science apps and let the individual programs handle it whatever way they want (like CPDN does), but then the Windows users, the majority, would complain about getting lower credits. So the BOINC Project eventually decided that the individual projects should just compensate for the lower credits on other platforms somehow, which most were already doing. This is an extremely simplified version and leaves out the high drama, tantrums and similar that occurred at other projects using BOINC. [Edit 6 times, last edit by Former Member at Nov 27, 2008 3:36:16 PM] |
||
|
|
Former Member
Cruncher Joined: May 22, 2018 Post Count: 0 Status: Offline |
Thank you for clearifying this.
However I can not agree in the following statement: Basically it is caused by the compiler used under Windows over-optimising the benchmark code in the BOINC client, while the one used under Linux and all other platforms doesn't. So actually the Linux results are accurate but the Windows ones are inflated. On another machine I get under Windows with Boinc 1538 floating point MIPS and 3325 integer MIPS per CPU. So because it is a DualCore CPU that would be 3076 floating point MIPS and 6650 integer MIPS. With SiSoft Sandra I get 10456 floating point MIPS and 13993 integer MIPS. So Sandra is 2-3 times faster than BOINC. So Boinc reports much weaker results than the CPUs are capable. If the software running under BOINC is also so weakly optimized then there is a lot of computation power wasted. Don't know about the optimization grade of the applications running under boinc. |
||
|
|
Former Member
Cruncher Joined: May 22, 2018 Post Count: 0 Status: Offline |
Hello Strom_umsonst,
With SiSoft Sandra I get 10456 floating point MIPS This is an incredible Whetstone figure. Just how overclocked is that CPU? Lawrence |
||
|
|
Sekerob
Ace Cruncher Joined: Jul 24, 2005 Post Count: 20043 Status: Offline |
The benchmark is utterly and completely irrelevant to the science processing and it's level of optimization. The only thing this test does is assess the device power and based on that attempts to compute a credit claim plus it's used to compute how much work to buffer. That's more or less it.
----------------------------------------oh, and since all devices say of windows measure the same way it helps [far from perfect] to establish homogeneity in the claims within a platform.
WCG
----------------------------------------Please help to make the Forums an enjoyable experience for All! [Edit 1 times, last edit by Sekerob at Dec 2, 2008 5:00:32 PM] |
||
|
|
Former Member
Cruncher Joined: May 22, 2018 Post Count: 0 Status: Offline |
This is an incredible Whetstone figure. Just how overclocked is that CPU? That CPU isn't overclocked at all. Its a Core 2 Duo Processor with 1,66 GHz. As I said. SiSoftSandra seams to get a lot of performance out of the processor. |
||
|
|
Former Member
Cruncher Joined: May 22, 2018 Post Count: 0 Status: Offline |
Strom_umsonst, SiSoft just use a different definition of MIP. You really can't compare different raw benchmark figures, or you will end up very, very frustrated.
|
||
|
|
widdershins
Veteran Cruncher Scotland Joined: Apr 30, 2007 Post Count: 677 Status: Offline Project Badges: 
|
Perhaps the best way to settle this would be for each WCG project to have benchmark units in a seperate project called "benchmark project". Say 50 units for each project in WCG. These same units would then be served up to every computer that joined the benchmark project and then a table could be produced showing performance of each main type of computer/os combination for each WU.
It could even be arranged so that "My Grid" showed each users PC performance compared with the average for similar PC's for each type of project. That way users could optimise their crunching by selecting the projects that suited their pc best and skipping ones experience shows they will perform badly on. This would also maxamise efficiency for WCG by ensuring the best tool was used for each WU I mainly crunch on 64 bit Linux as I find it more reliable (no BSOD's). I find it likes HCC units best, which happily coincides with my main crunching preferences. |
||
|
|
|