Finding The Perfect PC Components For Your Favorite Game Or Workload

[Michael]

The price is a significant factor for most of pc builders. I feel that there are few linux kernel developers who have 2100 usd cpu. Money is no subject for some companies and wintel religious gamers.

Right, just a matter of I need to update my Amazon API usage and such to get it accurately working again.

[Michael]

Awesome Michael.
Some Feedback:
- The ultimate quality of xonotic shows two times here but with totally different results: http://openbenchmarking.org/showdown/pts/xonotic

That's due to different versions of the test profile, e.g. when the version of Xonotic changes or other changes to the test are made that could impact the performance, the test profile major version is bumped to reflect that. So, yes, two versions are expected.

- It is not clear to me if one bar shows only one result of one system or of many different systems with the same graphics card. If one bar shows more than one system it would be nice to know how many systems are included in one bar - e.g. show the count as n=? next to SE +/-

They all show many systems... If there is like less than 10 or 20 systems with that given component, OpenBenchmarking.org doesn't even look at the component when graphing these results.

- It would be nice to optionally group some results to reduce the amount of bars, like group all HD 7950 results of 2015 to one bar. Of course, just grouping all HD7950 results of all time would make no sense because of driver developement.
- If grouping is too complicated maybe just filtering by date would be sufficient

So far haven't come up with a perfect grouping algorithm to the always properly detect the AIB of a graphics card, some graphics cards strings are slightly different, etc.

[brauliobo]

Michael, something it was confusing the first time I benchmarked using phoronix-test-suite was the description. Now I see I'm not the only user confused with that.

When the default description is used, that is, the detected and relevant associated hardware with the test (CPU/GPU/memory/disk/etc), then the labels on the graph makes complete sense.

To ilustrate the problem, see this link: http://openbenchmarking.org/result/1...HA-GPUBENCHM05
From the labels "Karini just after ubuntu install", "gpu-test" and others I can't infer the hardware and compare.

IMHO, that label should always be automatically guessed to avoid such confusions.

[chrisb]

Looks good. The CPU graphs don't seem to differentiate between number of processors? Would be interesting to see uniprocessor vs SMP where it makes a difference.

[chrisb]

The price is a significant factor for most of pc builders. I feel that there are few linux kernel developers who have 2100 usd cpu. Money is no subject for some companies and wintel religious gamers.

If you're just looking for raw CPU power there are some bargain older Xeons on EBay e.g. X5670 can be had for $100 and is one of the top processors in compile benchmarks. X5570 can be had for $25 and a pair will do a kernel compile in one minute.

[chrisb]

How come the 5960X result is so bad on the Linux kernel compile? It should be around 45 seconds.

[Michael]

How come the 5960X result is so bad on the Linux kernel compile? It should be around 45 seconds.

Depends on many factors, but I presume on many 5960X systems if the performance was poor for compile Linux kernel, the disks backing those systems weren't too fast.

[trifud]

There are faster GPUs on OpenBenchmarking that for some reason haven't made it in this comparison.

[chrisb]

Maybe also separate out whether or not the system is a VM or bare metal, it will make a big difference since there are people benchmarking shared VPS providers. e.g. this result I would guess is being counted as a E5-2680 which should be fast (it's 8 cores SNB CPU with 20MB cache) but it's obviously running under VMware with only 2 cores which invalidates the results (kernel compile is 357 seconds there). Whereas this is the same CPU but 2x with 32 cores and taking 78s. So it would make sense to consider showing separate results for factors like #cpus/#cores/bare metal/VM.

Misidentified CPU in kernel compile v4.3 - "Intel" at top (43 seconds)?

Are all of the results being shown? On 3.1 kernel compile X5570 is 422s but checking X5570 results shows only 5 kernel compiles on that CPU with values of 62,62,62,170,174?

Is there a graph showing all the results of a test on a particular CPU? Like all X5570 results for kernel compile?

[chrisb]

Depends on many factors, but I presume on many 5960X systems if the performance was poor for compile Linux kernel, the disks backing those systems weren't too fast.

Yes slow disk vs SSD would account for maybe 15ish seconds according to this comparison but I would've thought people paying $1000+ for a CPU would invest in a decent SSD.

Does PTS drop the disk caches before running a test? It should be doing that for reproducibility otherwise subsequent test runs are barely going to hit the disk.