Message ID | 87395ne9by.fsf@elfo.mitica |
---|---|
State | New |
Headers | show |
On Fri, Jun 22, 2012 at 6:59 AM, Juan Quintela <quintela@redhat.com> wrote: > > Hi > > As discussed on this week kvm call, here are the migration tree with > XBRLE + huge memory fixes integrated. There were a bug on huge memory > (corrected on the v3 version just sent) and another on XZBRLE patches. > Orit is going to resend patches addressing comments, but here the fix is > integrated already. > > How to test? > > Compile qemu with trace support: > > ./configure --enable-trace-backend=simple > > Create an event file: > > $ cat /tmp/events > savevm_section_start > savevm_section_end > $ > > Now launch qemu with normal command line: > > qemu ..... --trace events=/tmp/events > > And when execution finish, you can use a script like the following to > see the stalls that happen: > > $ cat ~/counter_simpletrace > $1 ~ /savevm_section_end/ { > /* Print savevm_section_stop line when >100 ms duration */ > if ($2 > 50000) { > printf("%s times_missing=%u\n", $0, times_missing++); > } > } > $ > > And now you can analize: > > ./scripts/simpletrace.py ./trace-events trace-32554 | awk -f > ~/counter_simpletrace > > (adjust paths as required) > > Notice that this are stalls on the iothread. vcpus stalls can still be > bigger. The number that is tested here is 50ms, so I get cases that are > a bit bigger (between 60-100ms with one 8GB guest that is dirtying 512MB > in a loop. > > If you get stalls, or drop connections, please let me know the load. > > Thanks, Juan. > > > The following changes since commit > 47ecbdf07ed2c37bdfd2d77137d01bb319ce13da: > > libcacard: build fixes (2012-06-21 20:04:24 +0000) > > are available in the git repository at: > > http://repo.or.cz/r/qemu/quintela.git migration-next > > for you to fetch changes up to 0844b46b5d7d8d6fa01134d170b3c48b7f32eab7: > > Add XBZRLE statistics (2012-06-22 15:24:15 +0200) > Currently this XBZRLE related stats are only available during the active migration but not after the completion. Can we get this information even after the completion of the XBZRLE migration (via info migrate) ? It would help to keep track of the final list of cache misses/hits etc. To the generic migration statistics that gets displayed via "info migrate" can you pl. consider adding the duration of Stage3 (i.e. "Downtime") as measured on the source host (yes there is probably some time spent in the resuming of the guest on the destination host...). Ideally this "Downtime" should be the pretty close to the pre-defined downtime (i.e. default value(30ms?) or whatever the user set it to via migrate_set_downtime command). But it looks like on larger guests that is not the case... Still reviewing the code and testing it out... Thanks Vinod > > ---------------------------------------------------------------- > Isaku Yamahata (1): > Add MigrationParams structure > > Juan Quintela (8): > Add spent time for migration > Add tracepoints for savevm section start/end > No need to iterate if we already are over the limit > Only TCG needs TLB handling > Only calculate expected_time for stage 2 > Exit loop if we have been there too long > dirty bitmap: abstract its use > Maintain the number of dirty pages > > Orit Wasserman (13): > Add missing check for host_from_stream_offset return value for > RAM_SAVE_FLAG_PAGE > Add migration capabilites > Add XBZRLE documentation > Add cache handling functions > Add uleb encoding/decoding functions > Add save_block_hdr function > Add debugging infrastructure > Change ram_save_block to return -1 if there are no more changes > Add migration_end function > Add xbzrle_encode_buffer and xbzrle_decode_buffer functions > Add XBZRLE to ram_save_block and ram_save_live > Add set_cachesize command > Add XBZRLE statistics > > Makefile.objs | 1 + > arch_init.c | 378 > ++++++++++++++++++++++++++++++++++++++++++++------ > block-migration.c | 8 +- > cache.c | 219 +++++++++++++++++++++++++++++ > cpu-all.h | 1 + > cutils.c | 29 ++++ > docs/xbzrle.txt | 115 +++++++++++++++ > exec-obsolete.h | 50 ++++--- > exec.c | 34 +++-- > hmp-commands.hx | 34 +++++ > hmp.c | 90 ++++++++++++ > hmp.h | 3 + > include/qemu/cache.h | 81 +++++++++++ > migration.c | 131 +++++++++++++++-- > migration.h | 30 +++- > monitor.c | 7 + > qapi-schema.json | 97 ++++++++++++- > qemu-common.h | 19 +++ > qmp-commands.hx | 99 +++++++++++++ > savevm.c | 115 ++++++++++++++- > sysemu.h | 3 +- > trace-events | 5 + > vmstate.h | 2 +- > 23 files changed, 1449 insertions(+), 102 deletions(-) > create mode 100644 cache.c > create mode 100644 docs/xbzrle.txt > create mode 100644 include/qemu/cache.h >
On 06/22/2012 11:43 PM, Vinod Chegu wrote: > > > On Fri, Jun 22, 2012 at 6:59 AM, Juan Quintela <quintela@redhat.com <mailto:quintela@redhat.com>> wrote: > > > Hi > > As discussed on this week kvm call, here are the migration tree with > XBRLE + huge memory fixes integrated. There were a bug on huge memory > (corrected on the v3 version just sent) and another on XZBRLE patches. > Orit is going to resend patches addressing comments, but here the fix is > integrated already. > > How to test? > > Compile qemu with trace support: > > ./configure --enable-trace-backend=simple > > Create an event file: > > $ cat /tmp/events > savevm_section_start > savevm_section_end > $ > > Now launch qemu with normal command line: > > qemu ..... --trace events=/tmp/events > > And when execution finish, you can use a script like the following to > see the stalls that happen: > > $ cat ~/counter_simpletrace > $1 ~ /savevm_section_end/ { > /* Print savevm_section_stop line when >100 ms duration */ > if ($2 > 50000) { > printf("%s times_missing=%u\n", $0, times_missing++); > } > } > $ > > And now you can analize: > > ./scripts/simpletrace.py ./trace-events trace-32554 | awk -f ~/counter_simpletrace > > (adjust paths as required) > > Notice that this are stalls on the iothread. vcpus stalls can still be > bigger. The number that is tested here is 50ms, so I get cases that are > a bit bigger (between 60-100ms with one 8GB guest that is dirtying 512MB > in a loop. > > If you get stalls, or drop connections, please let me know the load. > > Thanks, Juan. > > > The following changes since commit 47ecbdf07ed2c37bdfd2d77137d01bb319ce13da: > > libcacard: build fixes (2012-06-21 20:04:24 +0000) > > are available in the git repository at: > > http://repo.or.cz/r/qemu/quintela.git migration-next > > for you to fetch changes up to 0844b46b5d7d8d6fa01134d170b3c48b7f32eab7: > > Add XBZRLE statistics (2012-06-22 15:24:15 +0200) > > > Currently this XBZRLE related stats are only available during the active migration but not > after the completion. Can we get this information even after the completion of the XBZRLE migration (via info migrate) ? It would help to keep track of the final list of cache misses/hits etc. > Sure , I will add it in the next patch set. > To the generic migration statistics that gets displayed via "info migrate" can you pl. consider adding the duration of Stage3 (i.e. "Downtime") as measured on the source host (yes there is probably some time spent in the resuming of the guest on the destination host...). > > Ideally this "Downtime" should be the pretty close to the pre-defined downtime (i.e. default value(30ms?) or whatever the user set it to via migrate_set_downtime command). But it looks like on larger guests that is not the case... > > Still reviewing the code and testing it out... > > Thanks > Vinod > > > > > ---------------------------------------------------------------- > Isaku Yamahata (1): > Add MigrationParams structure > > Juan Quintela (8): > Add spent time for migration > Add tracepoints for savevm section start/end > No need to iterate if we already are over the limit > Only TCG needs TLB handling > Only calculate expected_time for stage 2 > Exit loop if we have been there too long > dirty bitmap: abstract its use > Maintain the number of dirty pages > > Orit Wasserman (13): > Add missing check for host_from_stream_offset return value for RAM_SAVE_FLAG_PAGE > Add migration capabilites > Add XBZRLE documentation > Add cache handling functions > Add uleb encoding/decoding functions > Add save_block_hdr function > Add debugging infrastructure > Change ram_save_block to return -1 if there are no more changes > Add migration_end function > Add xbzrle_encode_buffer and xbzrle_decode_buffer functions > Add XBZRLE to ram_save_block and ram_save_live > Add set_cachesize command > Add XBZRLE statistics > > Makefile.objs | 1 + > arch_init.c | 378 ++++++++++++++++++++++++++++++++++++++++++++------ > block-migration.c | 8 +- > cache.c | 219 +++++++++++++++++++++++++++++ > cpu-all.h | 1 + > cutils.c | 29 ++++ > docs/xbzrle.txt | 115 +++++++++++++++ > exec-obsolete.h | 50 ++++--- > exec.c | 34 +++-- > hmp-commands.hx | 34 +++++ > hmp.c | 90 ++++++++++++ > hmp.h | 3 + > include/qemu/cache.h | 81 +++++++++++ > migration.c | 131 +++++++++++++++-- > migration.h | 30 +++- > monitor.c | 7 + > qapi-schema.json | 97 ++++++++++++- > qemu-common.h | 19 +++ > qmp-commands.hx | 99 +++++++++++++ > savevm.c | 115 ++++++++++++++- > sysemu.h | 3 +- > trace-events | 5 + > vmstate.h | 2 +- > 23 files changed, 1449 insertions(+), 102 deletions(-) > create mode 100644 cache.c > create mode 100644 docs/xbzrle.txt > create mode 100644 include/qemu/cache.h > >