Re: [libvirt PATCH] ci: maximize cirrus CI ram/cpu allocation
On Wed, Jul 21, 2021 at 12:06:35PM +0100, Daniel P. Berrangé wrote: > On Wed, Jul 21, 2021 at 11:48:17AM +0100, Daniel P. Berrangé wrote: > > Thus I'm really puzzelled why libvirt didn't see a real benefit > > when I tested it. As you say though, FreeBSD builds are no worse > > than Linux builds, so it isn't critical for us. > > I guess looking at the CPU chart in this run explains it: > > https://cirrus-ci.com/task/5496170802839552 > > The first 75% of the job time was using less than 1 CPU. Only > the last 3 minutes maxed out both CPUs. So adding more CPUS > will only help the last bit of the job and even then we'll > hit limits on amount of parallelism due to build dependancies. > > QEMU simply has much much more stuff needing building and is > highly parallel when building, so saw a big benefit. > > So this affirms that we dont really need this patch. We could still spell out our expectations in term of resource allocation, and just set 4 CPUs / 2 GiB of RAM across the board. I think that would be still moderately useful, for documentation purposes if nothing else, and might help Cirrus CI better distribute its computing resources among users. We don't even need to make it dynamic per job, we can just hardcode those values in ci/cirrus/build.yml and call it a day. -- Andrea Bolognani / Red Hat / Virtualization
Re: [libvirt PATCH] ci: maximize cirrus CI ram/cpu allocation
On Wed, Jul 21, 2021 at 01:07:04PM +0200, Michal Prívozník wrote: > On 7/20/21 5:05 PM, Daniel P. Berrangé wrote: > > For macOS you always get the maximum configuration by default (12 CPUs, > > 24 GB RAM), but for FreeBSD you get 2 CPUs, 4 GBs by default. This > > change increases the allocation to 8 CPUs, 8 GBs for FreeBSD. > > > > Signed-off-by: Daniel P. Berrangé > > --- > > > > In theory this could make builds quicker. In practice I've not been > > able to measure a difference due to large variance between runs. > > > > .gitlab-ci.yml | 8 > > ci/cirrus/build.yml | 2 ++ > > 2 files changed, 10 insertions(+) > > A slightly offtopic question - is there a limit how many builds can we > run in given time frame, or how much resources can we consume? If so, > how far are we from getting rate limited? My understanding is that we > are using a free plan and I bet there would be some sort of limit. Cirrus doesn't limit overall consumption of resources, but it does limit job parallelism. IIRC, you can only have 2 jobs runing at any point in time in Cirrus. So we need to keep the number of jobs low. Regards, Daniel -- |: https://berrange.com -o-https://www.flickr.com/photos/dberrange :| |: https://libvirt.org -o-https://fstop138.berrange.com :| |: https://entangle-photo.org-o-https://www.instagram.com/dberrange :|
Re: [libvirt PATCH] ci: maximize cirrus CI ram/cpu allocation
On 7/20/21 5:05 PM, Daniel P. Berrangé wrote: > For macOS you always get the maximum configuration by default (12 CPUs, > 24 GB RAM), but for FreeBSD you get 2 CPUs, 4 GBs by default. This > change increases the allocation to 8 CPUs, 8 GBs for FreeBSD. > > Signed-off-by: Daniel P. Berrangé > --- > > In theory this could make builds quicker. In practice I've not been > able to measure a difference due to large variance between runs. > > .gitlab-ci.yml | 8 > ci/cirrus/build.yml | 2 ++ > 2 files changed, 10 insertions(+) A slightly offtopic question - is there a limit how many builds can we run in given time frame, or how much resources can we consume? If so, how far are we from getting rate limited? My understanding is that we are using a free plan and I bet there would be some sort of limit. Michal
Re: [libvirt PATCH] ci: maximize cirrus CI ram/cpu allocation
On Wed, Jul 21, 2021 at 11:48:17AM +0100, Daniel P. Berrangé wrote: > On Wed, Jul 21, 2021 at 06:39:07AM -0400, Andrea Bolognani wrote: > > On Tue, Jul 20, 2021 at 04:05:22PM +0100, Daniel P. Berrangé wrote: > > > For macOS you always get the maximum configuration by default (12 CPUs, > > > 24 GB RAM), but for FreeBSD you get 2 CPUs, 4 GBs by default. This > > > change increases the allocation to 8 CPUs, 8 GBs for FreeBSD. > > > > > > --- > > > In theory this could make builds quicker. In practice I've not been > > > able to measure a difference due to large variance between runs. > > > > Should we actually do this? The fact that FreeBSD builds take so much > > time is painful, but if you haven't been able to measure significant > > improvements from maximizing the number of CPUs assigned to the VM > > then perhaps other factors such as slow disk speed are to blame > > instead. In fact, looking at recent jobs it looks like they mostly > > fall in line with Linux builds... > > Yeah, I'm in two minds and just sent this patch to get a second > opinion. > > I noticed the ability to configure this when switching QEMU to > use lcitool for generating the freebsd/macos configuration. > The increase in resources for FreeBSD VMs made a significant > difference to QEMU build times - the difference between passing > and hitting the job limit timeout. > > Thus I'm really puzzelled why libvirt didn't see a real benefit > when I tested it. As you say though, FreeBSD builds are no worse > than Linux builds, so it isn't critical for us. I guess looking at the CPU chart in this run explains it: https://cirrus-ci.com/task/5496170802839552 The first 75% of the job time was using less than 1 CPU. Only the last 3 minutes maxed out both CPUs. So adding more CPUS will only help the last bit of the job and even then we'll hit limits on amount of parallelism due to build dependancies. QEMU simply has much much more stuff needing building and is highly parallel when building, so saw a big benefit. So this affirms that we dont really need this patch. Regards, Daniel -- |: https://berrange.com -o-https://www.flickr.com/photos/dberrange :| |: https://libvirt.org -o-https://fstop138.berrange.com :| |: https://entangle-photo.org-o-https://www.instagram.com/dberrange :|
Re: [libvirt PATCH] ci: maximize cirrus CI ram/cpu allocation
On Wed, Jul 21, 2021 at 06:39:07AM -0400, Andrea Bolognani wrote: > On Tue, Jul 20, 2021 at 04:05:22PM +0100, Daniel P. Berrangé wrote: > > For macOS you always get the maximum configuration by default (12 CPUs, > > 24 GB RAM), but for FreeBSD you get 2 CPUs, 4 GBs by default. This > > change increases the allocation to 8 CPUs, 8 GBs for FreeBSD. > > > > --- > > In theory this could make builds quicker. In practice I've not been > > able to measure a difference due to large variance between runs. > > Should we actually do this? The fact that FreeBSD builds take so much > time is painful, but if you haven't been able to measure significant > improvements from maximizing the number of CPUs assigned to the VM > then perhaps other factors such as slow disk speed are to blame > instead. In fact, looking at recent jobs it looks like they mostly > fall in line with Linux builds... Yeah, I'm in two minds and just sent this patch to get a second opinion. I noticed the ability to configure this when switching QEMU to use lcitool for generating the freebsd/macos configuration. The increase in resources for FreeBSD VMs made a significant difference to QEMU build times - the difference between passing and hitting the job limit timeout. Thus I'm really puzzelled why libvirt didn't see a real benefit when I tested it. As you say though, FreeBSD builds are no worse than Linux builds, so it isn't critical for us. > We certainly don't need 24 GiB of RAM to build libvirt, and grabbing > 12 CPUs seems excessively greedy too. Can we try being nice citizens > instead, and allocate something like 4 CPUs and 2 GiB of RAM across > the board? I expect that the resulting build times would still be > close to what we're seeing for containerized Linux builds, and so the > overall pipeline completion time won't be negatively affected. FWIW, for macOS systems you get the maximum resources no matter what, as they always allocate you an entire system. I presume this is because macOS builders run on physical hardware due to macOS license rules preventing VM usages. Regards, Daniel -- |: https://berrange.com -o-https://www.flickr.com/photos/dberrange :| |: https://libvirt.org -o-https://fstop138.berrange.com :| |: https://entangle-photo.org-o-https://www.instagram.com/dberrange :|
Re: [libvirt PATCH] ci: maximize cirrus CI ram/cpu allocation
On Tue, Jul 20, 2021 at 04:05:22PM +0100, Daniel P. Berrangé wrote: > For macOS you always get the maximum configuration by default (12 CPUs, > 24 GB RAM), but for FreeBSD you get 2 CPUs, 4 GBs by default. This > change increases the allocation to 8 CPUs, 8 GBs for FreeBSD. > > --- > In theory this could make builds quicker. In practice I've not been > able to measure a difference due to large variance between runs. Should we actually do this? The fact that FreeBSD builds take so much time is painful, but if you haven't been able to measure significant improvements from maximizing the number of CPUs assigned to the VM then perhaps other factors such as slow disk speed are to blame instead. In fact, looking at recent jobs it looks like they mostly fall in line with Linux builds... We certainly don't need 24 GiB of RAM to build libvirt, and grabbing 12 CPUs seems excessively greedy too. Can we try being nice citizens instead, and allocate something like 4 CPUs and 2 GiB of RAM across the board? I expect that the resulting build times would still be close to what we're seeing for containerized Linux builds, and so the overall pipeline completion time won't be negatively affected. -- Andrea Bolognani / Red Hat / Virtualization
