Re: [Qemu-devel] Fwd: [RFC 00/27] Migration thread (WIP)
Chegu Vinod chegu_vi...@hp.com wrote: On 7/26/2012 11:41 AM, Chegu Vinod wrote: Original Message Subject: [Qemu-devel] [RFC 00/27] Migration thread (WIP) Date: Tue, 24 Jul 2012 20:36:25 +0200 From: Juan Quintela quint...@redhat.com To: qemu-devel@nongnu.org Hi This series are on top of the migration-next-v5 series just posted. First of all, this is an RFC/Work in progress. Just a lot of people asked for it, and I would like review of the design. Hello, Thanks for sharing this early/WIP version for evaluation. Still in the middle of code review..but wanted to share a couple of quick observations. 'tried to use it to migrate a 128G/10VCPU guest (speed set to 10G and downtime 2s). Once with no workload (i.e. idle guest) and the second was with a SpecJBB running in the guest. The idle guest case seemed to migrate fine... capabilities: xbzrle: off Migration status: completed transferred ram: 3811345 kbytes remaining ram: 0 kbytes total ram: 134226368 kbytes total time: 199743 milliseconds In the case of the SpecJBB I ran into issues during stage 3...the source host's qemu and the guest hung. I need to debug this more... (if already have some hints pl. let me know.). capabilities: xbzrle: off Migration status: active transferred ram: 127618578 kbytes remaining ram: 2386832 kbytes total ram: 134226368 kbytes total time: 526139 milliseconds (qemu) qemu_savevm_state_complete called qemu_savevm_state_complete calling ram_save_complete --- hung somewhere after this ('need to get more info). Appears to be some race condition...as there are cases when it hangs and in some cases it succeeds. Weird guess, try to use less vcpus, same ram. The way that we stop cpus is _hacky_ to say it somewhere. Will try to think about that part. Thanks for the testing. All my testing has been done with 8GB guests and 2vcps. Will try with more vcpus to see if it makes a difference. (qemu) info migrate capabilities: xbzrle: off Migration status: completed transferred ram: 129937687 kbytes remaining ram: 0 kbytes total ram: 134226368 kbytes total time: 543228 milliseconds Humm, _that_ is more strange. This means that it finished. Could you run qemu under gdb and sent me the stack traces? I don't know your gdb thread kung-fu, so here are the instructions just in case: gdb --args exact qemu commandh line you used C-c to break when it hangs (gdb)info threads you see all the threads running (gdb)thread 1 or whatever other number (gdb)bt the backtrace of that thread. I am specially interested in the backtrace of the migration thread and of the iothread. Thanks, Juan. Need to review/debug... Vinod --- As with the non-migration-thread version the Specjbb workload completed before the migration attempted to move to stage 3 (i.e. didn't converge while the workload was still active). BTW, with this version of the bits (i.e. while running SpecJBB which is supposed to dirty quite a bit of memory) I noticed that there wasn't much change in the b/w usage of the dedicated 10Gb private network link (It was still ~1.5-3.0Gb/sec). Expected this to be a little better since we have a separate thread... not sure what else is in play here ? (numa locality of where the migration thread runs or something other basic tuning in the implementation ?) 'have a hi-level design question... (perhaps folks have already thought about it..and categorized it as potential future optimization..?) Would it be possible to off load the iothread completely [from all migration related activity] and have one thread (with the appropriate protection) get involved with getting the list of the dirty pages ? Have one or more threads dedicated for trying to push multiple streams of data to saturate the allocated network bandwidth ? This may help in large + busy guests. Comments? There are perhaps other implications of doing all of this (like burning more host cpu cycles) but perhaps this can be configurable based on user's needs... e.g. fewer but large guests on a host with no over subscription. Thanks Vinod
Re: [Qemu-devel] Fwd: [RFC 00/27] Migration thread (WIP)
Original Message Subject:[Qemu-devel] [RFC 00/27] Migration thread (WIP) Date: Tue, 24 Jul 2012 20:36:25 +0200 From: Juan Quintela quint...@redhat.com To: qemu-devel@nongnu.org Hi This series are on top of the migration-next-v5 series just posted. First of all, this is an RFC/Work in progress. Just a lot of people asked for it, and I would like review of the design. Hello, Thanks for sharing this early/WIP version for evaluation. Still in the middle of code review..but wanted to share a couple of quick observations. 'tried to use it to migrate a 128G/10VCPU guest (speed set to 10G and downtime 2s). Once with no workload (i.e. idle guest) and the second was with a SpecJBB running in the guest. The idle guest case seemed to migrate fine... capabilities: xbzrle: off Migration status: completed transferred ram: 3811345 kbytes remaining ram: 0 kbytes total ram: 134226368 kbytes total time: 199743 milliseconds In the case of the SpecJBB I ran into issues during stage 3...the source host's qemu and the guest hung. I need to debug this more... (if already have some hints pl. let me know.). capabilities: xbzrle: off Migration status: active transferred ram: 127618578 kbytes remaining ram: 2386832 kbytes total ram: 134226368 kbytes total time: 526139 milliseconds (qemu) qemu_savevm_state_complete called qemu_savevm_state_complete calling ram_save_complete --- hung somewhere after this ('need to get more info). --- As with the non-migration-thread version the Specjbb workload completed before the migration attempted to move to stage 3 (i.e. didn't converge while the workload was still active). BTW, with this version of the bits (i.e. while running SpecJBB which is supposed to dirty quite a bit of memory) I noticed that there wasn't much change in the b/w usage of the dedicated 10Gb private network link (It was still ~1.5-3.0Gb/sec). Expected this to be a little better since we have a separate thread... not sure what else is in play here ? (numa locality of where the migration thread runs or something other basic tuning in the implementation ?) 'have a hi-level design question... (perhaps folks have already thought about it..and categorized it as potential future optimization..?) Would it be possible to off load the iothread completely [from all migration related activity] and have one thread (with the appropriate protection) get involved with getting the list of the dirty pages ? Have one or more threads dedicated for trying to push multiple streams of data to saturate the allocated network bandwidth ? This may help in large + busy guests. Comments?There are perhaps other implications of doing all of this (like burning more host cpu cycles) but perhaps this can be configurable based on user's needs... e.g. fewer but large guests on a host with no over subscription. Thanks Vinod It does: - get a new bitmap for migration, and that bitmap uses 1 bit by page - it unfolds migration_buffered_file. Only one user existed. - it simplifies buffered_file a lot. - About the migration thread, special attention was giving to try to get the series review-able (reviewers would tell if I got it). Basic design: - we create a new thread instead of a timer function - we move all the migration work to that thread (but run everything except the waits with the iothread lock. - we move all the writting to outside the iothread lock. i.e. we walk the state with the iothread hold, and copy everything to one buffer. then we write that buffer to the sockets outside the iothread lock. - once here, we move to writting synchronously to the sockets. - this allows us to simplify quite a lot. And basically, that is it. Notice that we still do the iterate page walking with the iothread held. Light testing show that we got similar speed and latencies than without the thread (notice that almost no optimizations done here yet). Appart of the review: - Are there any locking issues that I have missed (I guess so) - stop all cpus correctly. vm_stop should be called from the iothread, I use the trick of using a bottom half to get that working correctly. but this _implementation_ is ugly as hell. Is there an easy way of doing it? - Do I really have to export last_ram_offset(), there is no other way of knowing the ammount of RAM? Known issues: - for some reason, when it has to start a 2nd round of bitmap handling, it decides to dirty all pages. Haven't found still why this happens. If you can test it, and said me where it breaks, it would also help. Work is based on Umesh thread work, and work that Paolo Bonzini had work on top of that. All the mirgation thread was done from scratch becase I was unable to debug why it was failing, but it owes a lot to the previous design. Thanks in advance, Juan. The following changes since commit a21143486b9c6d7a50b7b62877c02b3c686943cb: Merge remote-tracking branch
Re: [Qemu-devel] Fwd: [RFC 00/27] Migration thread (WIP)
On 7/26/2012 11:41 AM, Chegu Vinod wrote: Original Message Subject:[Qemu-devel] [RFC 00/27] Migration thread (WIP) Date: Tue, 24 Jul 2012 20:36:25 +0200 From: Juan Quintela quint...@redhat.com To: qemu-devel@nongnu.org Hi This series are on top of the migration-next-v5 series just posted. First of all, this is an RFC/Work in progress. Just a lot of people asked for it, and I would like review of the design. Hello, Thanks for sharing this early/WIP version for evaluation. Still in the middle of code review..but wanted to share a couple of quick observations. 'tried to use it to migrate a 128G/10VCPU guest (speed set to 10G and downtime 2s). Once with no workload (i.e. idle guest) and the second was with a SpecJBB running in the guest. The idle guest case seemed to migrate fine... capabilities: xbzrle: off Migration status: completed transferred ram: 3811345 kbytes remaining ram: 0 kbytes total ram: 134226368 kbytes total time: 199743 milliseconds In the case of the SpecJBB I ran into issues during stage 3...the source host's qemu and the guest hung. I need to debug this more... (if already have some hints pl. let me know.). capabilities: xbzrle: off Migration status: active transferred ram: 127618578 kbytes remaining ram: 2386832 kbytes total ram: 134226368 kbytes total time: 526139 milliseconds (qemu) qemu_savevm_state_complete called qemu_savevm_state_complete calling ram_save_complete --- hung somewhere after this ('need to get more info). Appears to be some race condition...as there are cases when it hangs and in some cases it succeeds. (qemu) info migrate capabilities: xbzrle: off Migration status: completed transferred ram: 129937687 kbytes remaining ram: 0 kbytes total ram: 134226368 kbytes total time: 543228 milliseconds Need to review/debug... Vinod --- As with the non-migration-thread version the Specjbb workload completed before the migration attempted to move to stage 3 (i.e. didn't converge while the workload was still active). BTW, with this version of the bits (i.e. while running SpecJBB which is supposed to dirty quite a bit of memory) I noticed that there wasn't much change in the b/w usage of the dedicated 10Gb private network link (It was still ~1.5-3.0Gb/sec). Expected this to be a little better since we have a separate thread... not sure what else is in play here ? (numa locality of where the migration thread runs or something other basic tuning in the implementation ?) 'have a hi-level design question... (perhaps folks have already thought about it..and categorized it as potential future optimization..?) Would it be possible to off load the iothread completely [from all migration related activity] and have one thread (with the appropriate protection) get involved with getting the list of the dirty pages ? Have one or more threads dedicated for trying to push multiple streams of data to saturate the allocated network bandwidth ? This may help in large + busy guests. Comments? There are perhaps other implications of doing all of this (like burning more host cpu cycles) but perhaps this can be configurable based on user's needs... e.g. fewer but large guests on a host with no over subscription. Thanks Vinod It does: - get a new bitmap for migration, and that bitmap uses 1 bit by page - it unfolds migration_buffered_file. Only one user existed. - it simplifies buffered_file a lot. - About the migration thread, special attention was giving to try to get the series review-able (reviewers would tell if I got it). Basic design: - we create a new thread instead of a timer function - we move all the migration work to that thread (but run everything except the waits with the iothread lock. - we move all the writting to outside the iothread lock. i.e. we walk the state with the iothread hold, and copy everything to one buffer. then we write that buffer to the sockets outside the iothread lock. - once here, we move to writting synchronously to the sockets. - this allows us to simplify quite a lot. And basically, that is it. Notice that we still do the iterate page walking with the iothread held. Light testing show that we got similar speed and latencies than without the thread (notice that almost no optimizations done here yet). Appart of the review: - Are there any locking issues that I have missed (I guess so) - stop all cpus correctly. vm_stop should be called from the iothread, I use the trick of using a bottom half to get that working correctly. but this _implementation_ is ugly as hell. Is there an easy way of doing it? - Do I really have to export last_ram_offset(), there is no other way of knowing the ammount of RAM? Known issues: - for some reason, when it has to start a 2nd round of bitmap handling, it decides to dirty all pages. Haven't found still why this happens. If you can test it, and said me where it breaks, it would also