Re: [Qemu-devel] block migration and MAX_IN_FLIGHT_IO
Am 05.03.2018 um 15:52 schrieb Dr. David Alan Gilbert: > * Peter Lieven (p...@kamp.de) wrote: >> Am 05.03.2018 um 12:45 schrieb Stefan Hajnoczi: >>> On Thu, Feb 22, 2018 at 12:13:50PM +0100, Peter Lieven wrote: I stumbled across the MAX_INFLIGHT_IO field that was introduced in 2015 and was curious what was the reason to choose 512MB as readahead? The question is that I found that the source VM gets very unresponsive I/O wise while the initial 512MB are read and furthermore seems to stay unreasponsive if we choose a high migration speed and have a fast storage on the destination VM. In our environment I modified this value to 16MB which seems to work much smoother. I wonder if we should make this a user configurable value or define a different rate limit for the block transfer in bulk stage at least? >>> I don't know if benchmarks were run when choosing the value. From the >>> commit description it sounds like the main purpose was to limit the >>> amount of memory that can be consumed. >>> >>> 16 MB also fulfills that criteria :), but why is the source VM more >>> responsive with a lower value? >>> >>> Perhaps the issue is queue depth on the storage device - the block >>> migration code enqueues up to 512 MB worth of reads, and guest I/O has >>> to wait? >> That is my guess. Especially if the destination storage is faster we >> basically alsways have >> 512 I/Os in flight on the source storage. >> >> Does anyone mind if the reduce that value to 16MB or do we need a better >> mechanism? > We've got migration-parameters these days; you could connect it to one > of those fairly easily I think. > Try: grep -i 'cpu[-_]throttle[-_]initial' for an example of one that's > already there. > Then you can set it to whatever you like. I will have a look at this. Thank you, Peter > > Dave > >> Peter >> >> > -- > Dr. David Alan Gilbert / dgilb...@redhat.com / Manchester, UK
Re: [Qemu-devel] block migration and MAX_IN_FLIGHT_IO
* Peter Lieven (p...@kamp.de) wrote: > Am 05.03.2018 um 12:45 schrieb Stefan Hajnoczi: > > On Thu, Feb 22, 2018 at 12:13:50PM +0100, Peter Lieven wrote: > >> I stumbled across the MAX_INFLIGHT_IO field that was introduced in 2015 > >> and was curious what was the reason > >> to choose 512MB as readahead? The question is that I found that the source > >> VM gets very unresponsive I/O wise > >> while the initial 512MB are read and furthermore seems to stay > >> unreasponsive if we choose a high migration speed > >> and have a fast storage on the destination VM. > >> > >> In our environment I modified this value to 16MB which seems to work much > >> smoother. I wonder if we should make > >> this a user configurable value or define a different rate limit for the > >> block transfer in bulk stage at least? > > I don't know if benchmarks were run when choosing the value. From the > > commit description it sounds like the main purpose was to limit the > > amount of memory that can be consumed. > > > > 16 MB also fulfills that criteria :), but why is the source VM more > > responsive with a lower value? > > > > Perhaps the issue is queue depth on the storage device - the block > > migration code enqueues up to 512 MB worth of reads, and guest I/O has > > to wait? > > That is my guess. Especially if the destination storage is faster we > basically alsways have > 512 I/Os in flight on the source storage. > > Does anyone mind if the reduce that value to 16MB or do we need a better > mechanism? We've got migration-parameters these days; you could connect it to one of those fairly easily I think. Try: grep -i 'cpu[-_]throttle[-_]initial' for an example of one that's already there. Then you can set it to whatever you like. Dave > Peter > > -- Dr. David Alan Gilbert / dgilb...@redhat.com / Manchester, UK
Re: [Qemu-devel] block migration and MAX_IN_FLIGHT_IO
Am 05.03.2018 um 12:45 schrieb Stefan Hajnoczi: > On Thu, Feb 22, 2018 at 12:13:50PM +0100, Peter Lieven wrote: >> I stumbled across the MAX_INFLIGHT_IO field that was introduced in 2015 and >> was curious what was the reason >> to choose 512MB as readahead? The question is that I found that the source >> VM gets very unresponsive I/O wise >> while the initial 512MB are read and furthermore seems to stay unreasponsive >> if we choose a high migration speed >> and have a fast storage on the destination VM. >> >> In our environment I modified this value to 16MB which seems to work much >> smoother. I wonder if we should make >> this a user configurable value or define a different rate limit for the >> block transfer in bulk stage at least? > I don't know if benchmarks were run when choosing the value. From the > commit description it sounds like the main purpose was to limit the > amount of memory that can be consumed. > > 16 MB also fulfills that criteria :), but why is the source VM more > responsive with a lower value? > > Perhaps the issue is queue depth on the storage device - the block > migration code enqueues up to 512 MB worth of reads, and guest I/O has > to wait? That is my guess. Especially if the destination storage is faster we basically alsways have 512 I/Os in flight on the source storage. Does anyone mind if the reduce that value to 16MB or do we need a better mechanism? Peter
Re: [Qemu-devel] block migration and MAX_IN_FLIGHT_IO
On Thu, Feb 22, 2018 at 12:13:50PM +0100, Peter Lieven wrote: > I stumbled across the MAX_INFLIGHT_IO field that was introduced in 2015 and > was curious what was the reason > to choose 512MB as readahead? The question is that I found that the source VM > gets very unresponsive I/O wise > while the initial 512MB are read and furthermore seems to stay unreasponsive > if we choose a high migration speed > and have a fast storage on the destination VM. > > In our environment I modified this value to 16MB which seems to work much > smoother. I wonder if we should make > this a user configurable value or define a different rate limit for the block > transfer in bulk stage at least? I don't know if benchmarks were run when choosing the value. From the commit description it sounds like the main purpose was to limit the amount of memory that can be consumed. 16 MB also fulfills that criteria :), but why is the source VM more responsive with a lower value? Perhaps the issue is queue depth on the storage device - the block migration code enqueues up to 512 MB worth of reads, and guest I/O has to wait? Stefan signature.asc Description: PGP signature
[Qemu-devel] block migration and MAX_IN_FLIGHT_IO
Hi, I stumbled across the MAX_INFLIGHT_IO field that was introduced in 2015 and was curious what was the reason to choose 512MB as readahead? The question is that I found that the source VM gets very unresponsive I/O wise while the initial 512MB are read and furthermore seems to stay unreasponsive if we choose a high migration speed and have a fast storage on the destination VM. In our environment I modified this value to 16MB which seems to work much smoother. I wonder if we should make this a user configurable value or define a different rate limit for the block transfer in bulk stage at least? Peter