Something a little unusual?

[mwroggenbuck]

All,

I have been getting more of the a8 jobs recently. However, some of these jobs do not use the full amount of CPU allocated to it. They use the proper number of threads, but each thread is not at 100% utilization. Current, I am running project 16814 and it is using approximately 90% of allocated CPU. There is nothing wrong with this. It is just that I have never seen it before. When ever I looked before, all threads allocated to a CPU job would run at 100%. Of course, most of those jobs were a7 work units.

Has anyone else noticed this? Like I said, I don't think there is anything wrong (jobs still work relatively fast). It is just that I have not seen this before, and I have seen it at least twice (both with a8)

Mark

[Joe_H]

The A8 core was released with a compilation setting that does not tie the processing to specific cores as much as the A7. The developer who worked on it ran into a bug with different settings for that option and is looking into that. So each processing thread may move from one core to another more often.

Basically nothing is wrong, just the physical processing hardware is utilized a bit differently.

[bruce]

I suspect (but have not ascertained) that internally FAHCore_a8 does have some Domain-Decomposition issues similar to FAHCore_a7 but they're not as visible. For example, if I have a 12 thread CPU and one GPU, FAH allocates CPU:11 to FAHCore_a7. Then the server will reduce the number of active cores to avoid the DD errors and that's explicitly reported in the log. FAHCore_a8 is able to deal with 11 available cores, but not in the same way and it's all behind-the-scenes. Maybe that's why you see 90% rather than 99%.

[BobWilliams757]

I haven't paid much attention to the CPU use, but the A8 cores sure do run well. Understanding that for now the A8 PPD returns are still inflated, but they do seem to coexist better when both CPU and GPU folding on my APU system. It was almost not worth it with the A7 cores, as the CPU folding just slowed the GPU folding, and the net gain was very little. Now they seem to get along better and overall throughput increases almost by the CPU output as if it was running by itself.

I'm calling it a win either way on my system.

[PantherX]

I would be keen for some log files to see what your expectation is, what reality is and what F@H is doing. Considering that FahCore_a8 is under active development, if you do happen to stumble upon an edge case, it can hopefully be addressed sooner rather than later

[mwroggenbuck]

I don't really have any log files to show you. I am just using HWINFO64 (https://www.hwinfo.com/download/) to monitor the CPU load (and GPU load). Recently, the a8 core work units have been using less CPU. I suspect this is project related, because I seem to remember looking at a8 work units a while back and seeing they were taking all of CPU, and even more power (I suspect it has to do with avx2_256). Those work units were flat out stressing my CPU as much as they could.

Like I said earlier, their is nothing wrong as far as I can tell. It is just that I had never seen these jobs not use as much CPU as they have been allocated (except for when the client reduces the number of threads, which is not happening here). If you want to look for something, start with the projects (my current project, which is showing less CPU usage is 16946)

[mwroggenbuck]

Quick update:

I just received an a7 work unit (project 16460). It is definitely using all the CPU allocated to it.

[Knish]

my i5 4-core is with an a8 at 85% now with cpu:3 since 1 is for a gpu. Does this help with anyone's theories?

[mwroggenbuck]

Another a8 work unit (16943). Only using about 3/4 of available CPU!

[bruce]

FAH is designed to use all of the CPU resources that are devoted to it ... but at an extremely low priority. Since they're running at an extremely low priority, all you foreground activities will be running at a higher priority so FAH will use all UNUSED resources.

You can reduce the number of threads allocated to FAH in FAHControl + Configure

[mwroggenbuck]

Except that System Idle Process is using what is left over, so there is plenty of unused resources. For some reason, these work units simply do not stress the system. That is OK. I just never have seen it before. Maybe these work units hit memory more frequently, and leave the CPU idle during that time.

[bruce]

A GPU kernel will contiue to use the GPU until it finishes (assuming you're looking a GPU compute status). One CPU thread needs to be available to supply the GPU with the next kernel. FAH allocates on thread to each GPU but if you load up those CPUs with BOINC (or something else) it will slow down GPU processing.

Running both BOINC and FAH is not recommended. Choose one or the other or intentionally leave some CPU headroom. A modicum of idle time is needed.

[mwroggenbuck]

Only running FAH. GPU loads do not change if I run CPU slot or not, so I don't think GPU is affecting the CPU that I allocate to the CPU, and I always leave some some idle available (at least 10%). Over the last couple of days, I can definitely see the difference between a7 and a8 even though the FAH settings are the same in all cases. I have not bothered to monitor cpu states--it is possible that a8 has more wait states (spin locks, etc). The bottom line is that I see a7 stressing the CPU much more than a8. I was wondering if anyone else was seeing that. Maybe avx2_256 is more efficient than avx_256?

[bruce]

Yes, _a8 is able to utilize more SSE/AVX options than _a7 so that probably explains the increase heat from your CPU, depending on the features of your CPU.