So, I ran perf on my host and it came back far more true. The top consumers of time were all atomics and some function called sse3, which I believe is a super fast memcpy implementation provided the the arch. In addition, all the highest time consumers are within my image- it stayed out of the kernel as designed and it used additional extensions and features. I just thought of something-what if there is some kind of page size difference between my host and my Linux kernel causing the performance problems? On Nov 24, 2016 11:33 AM, "Kenneth Adam Miller" <kennethadammiller@gmail.com> wrote:
On Thu, Nov 24, 2016 at 11:13 AM, Greg KH <greg@kroah.com> wrote:
On Thu, Nov 24, 2016 at 10:31:18AM -0500, Kenneth Adam Miller wrote:
On Nov 24, 2016 2:18 AM, "Greg KH" <greg@kroah.com> wrote:
On Thu, Nov 24, 2016 at 02:01:41AM -0500, Kenneth Adam Miller wrote:
Hello,
I have a scheduler issue in two different respects:
1) I have a process that is supposed to tight loop, and it is being given very very little time on the system. I don't want that - I
want
those who would use the processor to be given the resources to run as fast as they each can.
What is causing it to give up its timeslice? Is it waiting for I/O? Doing something else to sleep?
It's multithreaded, so it reads in a loop in one thread and writes in another thread. What I saw when I ran strace on it is each process would run for too long- the program is designed to try and stay out of the kernel on each side, so it checks some shared variables before it ever goes.
So locking/cpu contention for those "shared variables" perhaps?
I don't think that could possibly be it, because the shared variables are controlled by atomics. It's just some memory operation to check to see if it needs to go to the kernel, as in is there more data in the shm region for me to read? If not, I'll go wait on this OS semaphore. It's lightening fast on my host machine.
2) I am seeing with perf that the maximum overhead at each section does not sum up to be more than 15 percent. Total, probably
something
like 18% of cpu time is used, and my binary has rocketed in slowness from about 2 seconds or less total to several minutes.
What changed to make things slower? Did you change kernel versions or did you change something in your userspace program?
The kernel versions specifically couldnt have anything to do with it but it was different kernels. The test runs in less that 2 seconds on my host. When I copy it to our custom linux, it takes minutes for it to run. I think it's some extra setting that we're missing while building the kernel, and I don't know what that is. I got a huge improvement when I changed the multicore scheduling to allow preemption "(desktop)" but there's still a problem as I've described with one of the processes not using the core as it should.
What do you mean by "custom linux"? Is this the exact same hardware as your machine? Or different? If so, what is different? What is different between the different kernel versions you are using? Does the perf output look different from running on the two different machines? If so, where?
I am building with buildroot a linux that is meant to be really stripped down and only have the things we want. In my case, the what the bzImage sees is either what QEMU gives it or what it sees in our dedicated hardware, with is just off the shelf i7 and other stuff you get a market - nothing custom in the sense you are thinking. Custom as in, roll your own linux.
The kernel versions between my host and the target are 3.13.x and 3.14.5x; they don't change so much, and certainly don't affect performance on their own. I'm missing some setting or something with how I'm configuring or building linux.
I haven't had a chance to run perf on my host. I can't find what ubuntu package it is just yet, but I will search for it in a minute. I have to go somewhere and will be right back immediately.
Have you changed the priority levels of your application at all? Have you thought about just forcing your app to a specific CPU and getting the kernel off of that CPU in order so that the kernel isn't even an option here at all (Linux allows you to do this, details are somewhere in the documentation, sorry, can't remember off the top of my head...)
No, that may be it or help though. I thought that binding an application to a particular cpu had something to do with affinity and that there was some C api for it or something. That would work for our particular scenario, and we've even talked about it, I just don't know how to do it yet.
But really, you should track down what the differences are between your two machines/environments, as something is different that is causing the slow down.
True - the kernel configuration is most suspect based on everything I know. The hardware differences between my host to the target we're building for is each modern, and well supported by linux. I'm thinking it absolutely must have something to do with the way I've built linux.
You haven't even said what kernel version you are using, and if you have any of your own kernel patches in those kernels.
For the target hardware is 3.14.5x, and there aren't any kernel patches at this time; I've disabled grsec while in the process of narrowing down what the problem is.
I think that the linux scheduler isn't scheduling it, because this process is just some unit tests that double as benchmarks in that they shm_open a file and write into it with memcpy's.
Are you sure that I/O isn't happening here like through swap or something else?
Well, we're using tmpfs and don't have a disk in the machine, but I will say this process is using all lot of the address space. One problem here is that the kernel has more ram than it thinks it does,
What do you mean, is this a hardware issue?
I don't think it's hardware; we're using this proprietary software beneath the linux kernel, but it's still ram of course. I can't say too too much, but what I can say is that while how much linux thinks it has could be affecting how it behaves, on our end we have the resources and can just change the configuration to make sure that linux sees and has enough ram. So that we can test on our end, and indeed we will.
but what I want to emphasize is that I haven't changed the program to allocate any more than it was previously. I'm not sure if that's a kernel change or some setting, but it went from 85% to 98%.
What exactly went up by 17%?
Consider the process that I was talking about that is meant to tight loop and burn on a core to be the "end product process". This is different from the test benchmarks that I was explaining run so poorly.
The reason why is that there is a large latency even without that big program in there; I can't run my standalone tests in qemu without it also taking minutes. I understand qemu has to emulate, and that's its not just a VM, but I'm going from host CPU to guest, and the settings are the same.
That doesn't really make much sense, why is qemu even in the picture here? And no, qemu doesn't always emulate things, that depends on the hardware you are running it on, and what type of image you are running on it.
Well, when I'm not at work, I have to be able to run the bzImage on something, and I don't have a dedicated machine. So I run it in QEMU.
What does perf say is taking all of your time?
When I ran perf what it appeared to indicate is that the largest consumer of time was my library, which should be right in either scenario because it should use stay out of the kernel as I've designed it. In addition, the work takes place there anyway, so that's right. What's not right is the fact that the largest percent of time used is around 15%, and all the others combined don't add up to anything near 100.
So perhaps you have other processes running on the machine that you are not noticing that is taking up the time slices? Are you _sure_ nothing else is running?
I'm certain that there are other processes alive, but they are not using the CPU. This process is the only one running. I even gave qemu "-smp 4" because I want it to behave as close as possible to what it would if it were just on the host.
Basically, you have a bunch of variables, and haven't been very specific with what really is changing, or even being used here, so there's not much specific that I can think of at the moment.
thanks,
greg k-h