Governor for Phoronix benchmarks

[matyas]

The recent articles pinpointed the fact that the use of different governor can have a huge impact on benchmarks. It would be interesting to see what Phoronix readers think about which governor should be used for benchmarks. And let's not make this a "Michael change the governor because of the pool thing" please.

[Vim_User]

It depends. If Michael is running a benchmark to look for improvements in CPU power-management it doesn't make much sense to use the performance governor. This should be handled on a case-by-case basis, if you measure something like CPU power-management use ondemand, if you test CPU performance in different scenarios use ondemand (AFAIK, using performance also disables features like Turbo Core that can improve single-thread performance), if you for example test video drivers make the testbed as invariable as possible (except the components to test, of course), so use performance here.

So I didn't vote here for either one.

[schmidtbag]

As I have stated in my response to the article, the governor Michael was using was fair. While testing for highest optimal results is absolutely interesting, the important results are the ones that the average user will be involved with.

I feel that the CPU should not have had as immense of an impact on the radeon GPU performance. Considering intel's GPU drivers were overall unaffected, this leads me to think there is something wrong with the radeon driver being so dependent upon CPU frequency - it doesn't appear to be too CPU intensive (hence the CPU downclocking itself all the time), but rather it seems to need a more "responsive" system.

As for anyone thinking that the performance governor is implied for benchmarks, I'd have to say that's a very naive thought. First of all, I'm not aware of any Windows benchmarks that do this. I'm sure there are plenty out there that do, but I get the impression most don't. You could argue that I'm comparing apples to oranges, but the point is the governor should never have this drastic of a negative impact on performance. That could mean intel's ondemand governor was faulty all this time, or, it could mean the radeon drivers are way too dependent upon CPU cycles. If Catalyst suffers the same problem, it could mean the Radeon hardware has issues.

The second problem with always using the performance governor is that means the GPU needs the same treatment - DPM should always be disabled and the GPU itself should always run maxed out. This is NOT realistic - people should not have their GPUs running at 70C when browsing the web. The third problem is Michael tests more than just the best performance - he also tests power efficiency. When you set the governor to performance, or disable DPM, you skew the results. If you argue "but they're on by default", well, then you're a hypocrite.

[Temar]

You are missing one option in your poll:

[x] The kernel's default scheduler

[s_j_newbury]

Also missing the intel p-state driver, which is the kernel default on recent Intel CPUs. (except for Ubuntu)

[s_j_newbury]

I feel that the CPU should not have had as immense of an impact on the radeon GPU performance. Considering intel's GPU drivers were overall unaffected, this leads me to think there is something wrong with the radeon driver being so dependent upon CPU frequency - it doesn't appear to be too CPU intensive (hence the CPU downclocking itself all the time), but rather it seems to need a more "responsive" system.

This is an interesting point. It could well be the driver has a commonly hit but (usually) fast GL CPU fallback or synchronous function call which has tight latency requirements. It's something to look for in performance profiling.

[matyas]

Should GPU drivers be in the business of driving the CPU frequency?

[bridgman]

I feel that the CPU should not have had as immense of an impact on the radeon GPU performance. Considering intel's GPU drivers were overall unaffected, this leads me to think there is something wrong with the radeon driver being so dependent upon CPU frequency - it doesn't appear to be too CPU intensive (hence the CPU downclocking itself all the time), but rather it seems to need a more "responsive" system.

Not sure this is a valid argument, given that the radeon performance difference primarily showed on high end GPUs, with little or no difference on GPUs in the same performance range as Intel GPU hardware. On smaller GPUs the performance tends to hit GPU limits first.

It's not as simple as just being GPU limited or CPU limited, of course, (more like the logic for parallel resistors) but at first glance only the high end GPUs can process enough graphics data for CPU throughput to become a significant factor.

[GreatEmerald]

Right now ondemand seems to be performance. So I don't really care. A much more interesting question would be whether it should be powersave or performance – when talking about the intel_pstate driver, of course.

[schmidtbag]

Not sure this is a valid argument, given that the radeon performance difference primarily showed on high end GPUs, with little or no difference on GPUs in the same performance range as Intel GPU hardware. On smaller GPUs the performance tends to hit GPU limits first.

It's not as simple as just being GPU limited or CPU limited, of course, (more like the logic for parallel resistors) but at first glance only the high end GPUs can process enough graphics data for CPU throughput to become a significant factor.

That is a very good point about the low-end hardware being much less impacted. As I have stated in the past, there is not enough data to be pointing fingers at anyone or anything, and we need tests with Catalyst to know whether or not this is a driver issue or a governor issue. Monitoring of CPU frequency for these tests is also important.

I think nvidia binary and nouveau tests would also help, since their hardware is better than intel's.


Regardless of what's really going on, I'm still grateful for the radeon driver developers and what they have accomplished.