| Message ID | 20201026171815.13233-21-vsementsov@virtuozzo.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | backup performance: block_status + async | expand |
On 26.10.20 18:18, Vladimir Sementsov-Ogievskiy wrote: > Further commit will add a benchmark > (scripts/simplebench/bench-backup.py), which will show that backup > works better with async parallel requests (previous commit) and > disabled copy_range. So, let's disable copy_range by default. > > Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> > --- > qapi/block-core.json | 2 +- > blockdev.c | 2 +- > 2 files changed, 2 insertions(+), 2 deletions(-) Uh, well, er, then why did you add it as true in patch 2? Do you mean this patch as the basis for a discussion whether it should be true or false by default? I don’t have anything to contribute, though, ergo I don’t mind either way. Do you have an idea why copy offloading isn’t better? Max > diff --git a/qapi/block-core.json b/qapi/block-core.json > index 5a21c24c1d..58eb2bcb86 100644 > --- a/qapi/block-core.json > +++ b/qapi/block-core.json > @@ -1376,7 +1376,7 @@ > # Optional parameters for backup. These parameters don't affect > # functionality, but may significantly affect performance. > # > -# @use-copy-range: Use copy offloading. Default true. > +# @use-copy-range: Use copy offloading. Default false. > # > # @max-workers: Maximum number of parallel requests for the sustained background > # copying process. Doesn't influence copy-before-write operations. > diff --git a/blockdev.c b/blockdev.c > index 0ed390abe0..1ac64d8ee2 100644 > --- a/blockdev.c > +++ b/blockdev.c > @@ -2788,7 +2788,7 @@ static BlockJob *do_backup_common(BackupCommon *backup, > { > BlockJob *job = NULL; > BdrvDirtyBitmap *bmap = NULL; > - BackupPerf perf = { .use_copy_range = true, .max_workers = 64 }; > + BackupPerf perf = { .max_workers = 64 }; > int job_flags = JOB_DEFAULT; > > if (!backup->has_speed) { >
12.01.2021 17:05, Max Reitz wrote: > On 26.10.20 18:18, Vladimir Sementsov-Ogievskiy wrote: >> Further commit will add a benchmark >> (scripts/simplebench/bench-backup.py), which will show that backup >> works better with async parallel requests (previous commit) and >> disabled copy_range. So, let's disable copy_range by default. >> >> Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> >> --- >> qapi/block-core.json | 2 +- >> blockdev.c | 2 +- >> 2 files changed, 2 insertions(+), 2 deletions(-) > > Uh, well, er, then why did you add it as true in patch 2? > > Do you mean this patch as the basis for a discussion whether it should be true or false by default? Yes, I just kept the change in separate, to make it simpler to discuss.. Hmm, or seems I've tried to keep it close to the point when we are going to increase performance anyway (and making default false is needed for good final performance), when if we do the change earlier it would like a degradation during previous 20 commits. > I don’t have anything to contribute, though, ergo I don’t mind either way. > > Do you have an idea why copy offloading isn’t better? > As I understand, it may be really fast for filesystem which will avoid data copying, like btrfs. Why it works slower for others when we enable async + block-status things, I don't know. It should depend on kernel implementation. I have two additional thoughts: 1. for backup we usually want to do a copy on another node, or at least another hard drive, so we really going to copy the data, not just create some metadata links in fs. 2. even if fs can just create some links to existing clusters instead of copying data, IMHO it is not good for backup: Main difference of backup from snapshot is that we don't touch active disk: it works in the same way like without backup job running. But if we do create fs-metadata links in backup image to original clusters, it means that on next guest write fs will have to do COW operation, so backup file will reference old cluster from active disk, and new cluster will be allocated for active disk. This means, that backup job influence on active vm disk, increasing its fragmentation.. Possibly I'm wrong, I'm not really good in how filesystems works. > >> diff --git a/qapi/block-core.json b/qapi/block-core.json >> index 5a21c24c1d..58eb2bcb86 100644 >> --- a/qapi/block-core.json >> +++ b/qapi/block-core.json >> @@ -1376,7 +1376,7 @@ >> # Optional parameters for backup. These parameters don't affect >> # functionality, but may significantly affect performance. >> # >> -# @use-copy-range: Use copy offloading. Default true. >> +# @use-copy-range: Use copy offloading. Default false. >> # >> # @max-workers: Maximum number of parallel requests for the sustained background >> # copying process. Doesn't influence copy-before-write operations. >> diff --git a/blockdev.c b/blockdev.c >> index 0ed390abe0..1ac64d8ee2 100644 >> --- a/blockdev.c >> +++ b/blockdev.c >> @@ -2788,7 +2788,7 @@ static BlockJob *do_backup_common(BackupCommon *backup, >> { >> BlockJob *job = NULL; >> BdrvDirtyBitmap *bmap = NULL; >> - BackupPerf perf = { .use_copy_range = true, .max_workers = 64 }; >> + BackupPerf perf = { .max_workers = 64 }; >> int job_flags = JOB_DEFAULT; >> if (!backup->has_speed) { >> >
On 12.01.21 16:44, Vladimir Sementsov-Ogievskiy wrote: > 12.01.2021 17:05, Max Reitz wrote: >> On 26.10.20 18:18, Vladimir Sementsov-Ogievskiy wrote: >>> Further commit will add a benchmark >>> (scripts/simplebench/bench-backup.py), which will show that backup >>> works better with async parallel requests (previous commit) and >>> disabled copy_range. So, let's disable copy_range by default. >>> >>> Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> >>> --- >>> qapi/block-core.json | 2 +- >>> blockdev.c | 2 +- >>> 2 files changed, 2 insertions(+), 2 deletions(-) >> >> Uh, well, er, then why did you add it as true in patch 2? >> >> Do you mean this patch as the basis for a discussion whether it should >> be true or false by default? > > Yes, I just kept the change in separate, to make it simpler to discuss.. > Hmm, or seems I've tried to keep it close to the point when we are going > to increase performance anyway (and making default false is needed for > good final performance), when if we do the change earlier it would like > a degradation during previous 20 commits. Hm, yes, before this series, copy_range is effectively the default, right. >> I don’t have anything to contribute, though, ergo I don’t mind either >> way. >> >> Do you have an idea why copy offloading isn’t better? >> > > As I understand, it may be really fast for filesystem which will avoid > data copying, like btrfs. Why it works slower for others when we enable > async + block-status things, I don't know. It should depend on kernel > implementation. > > I have two additional thoughts: > > 1. for backup we usually want to do a copy on another node, or at least > another hard drive, so we really going to copy the data, not just create > some metadata links in fs. > > 2. even if fs can just create some links to existing clusters instead of > copying data, IMHO it is not good for backup: > > Main difference of backup from snapshot is that we don't touch active > disk: it works in the same way like without backup job running. But if > we do create fs-metadata links in backup image to original clusters, it > means that on next guest write fs will have to do COW operation, so > backup file will reference old cluster from active disk, and new cluster > will be allocated for active disk. This means, that backup job influence > on active vm disk, increasing its fragmentation.. Possibly I'm wrong, > I'm not really good in how filesystems works. Both sound like good points to me. So as far as I’m concerned, disabling copy offloading sounds good to me. So now this separate patch here makes sense to me: Reviewed-by: Max Reitz <mreitz@redhat.com> Perhaps it deserves some explanation in a commit message, though. I.e., that originally, this defaulted to true, as it was before, but now that async copying is implemented, we should let it default to false. Max
diff --git a/qapi/block-core.json b/qapi/block-core.json index 5a21c24c1d..58eb2bcb86 100644 --- a/qapi/block-core.json +++ b/qapi/block-core.json @@ -1376,7 +1376,7 @@ # Optional parameters for backup. These parameters don't affect # functionality, but may significantly affect performance. # -# @use-copy-range: Use copy offloading. Default true. +# @use-copy-range: Use copy offloading. Default false. # # @max-workers: Maximum number of parallel requests for the sustained background # copying process. Doesn't influence copy-before-write operations. diff --git a/blockdev.c b/blockdev.c index 0ed390abe0..1ac64d8ee2 100644 --- a/blockdev.c +++ b/blockdev.c @@ -2788,7 +2788,7 @@ static BlockJob *do_backup_common(BackupCommon *backup, { BlockJob *job = NULL; BdrvDirtyBitmap *bmap = NULL; - BackupPerf perf = { .use_copy_range = true, .max_workers = 64 }; + BackupPerf perf = { .max_workers = 64 }; int job_flags = JOB_DEFAULT; if (!backup->has_speed) {
Further commit will add a benchmark (scripts/simplebench/bench-backup.py), which will show that backup works better with async parallel requests (previous commit) and disabled copy_range. So, let's disable copy_range by default. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> --- qapi/block-core.json | 2 +- blockdev.c | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-)