| Message ID | 1617938456-315058-10-git-send-email-lei.rao@intel.com |
|---|---|
| State | New |
| Headers | show |
| Series | Fixed some bugs and optimized some codes for COLO | expand |
Hi, Dave The performance data has added to the commit messages. Do you have any other suggestions? Thanks Lei. -----Original Message----- From: Rao, Lei <lei.rao@intel.com> Sent: Friday, April 9, 2021 11:21 AM To: Zhang, Chen <chen.zhang@intel.com>; lizhijian@cn.fujitsu.com; jasowang@redhat.com; quintela@redhat.com; dgilbert@redhat.com; pbonzini@redhat.com; lukasstraub2@web.de Cc: qemu-devel@nongnu.org; Rao, Lei <lei.rao@intel.com> Subject: [PATCH v6 09/10] Add the function of colo_bitmap_clear_dirty From: "Rao, Lei" <lei.rao@intel.com> When we use continuous dirty memory copy for flushing ram cache on secondary VM, we can also clean up the bitmap of contiguous dirty page memory. This also can reduce the VM stop time during checkpoint. The performance test for COLO as follow: Server configuraton: CPU :Intel(R) Xeon(R) Gold 6140 CPU @ 2.30GHz MEM :251G(type:DDR4 Speed:2666 MT/s) SSD :Intel 730 and DC S35x0/3610/3700 Series SSDs dirty pages:3189376 migration_bitmap_clear_dirty time consuming(ns):105194000 dirty pages:3189784 migration_bitmap_clear_dirty time consuming(ns):105297000 dirty pages:3190501 migration_bitmap_clear_dirty time consuming(ns):105410000 dirty pages:3188734 migration_bitmap_clear_dirty time consuming(ns):105138000 dirty pages:3189464 migration_bitmap_clear_dirty time consuming(ns):111736000 dirty pages:3188558 migration_bitmap_clear_dirty time consuming(ns):105079000 dirty pages:3239489 migration_bitmap_clear_dirty time consuming(ns):106761000 dirty pages:3190240 colo_bitmap_clear_dirty time consuming(ns):8369000 dirty pages:3189293 colo_bitmap_clear_dirty time consuming(ns):8388000 dirty pages:3189171 colo_bitmap_clear_dirty time consuming(ns):8641000 dirty pages:3189099 colo_bitmap_clear_dirty time consuming(ns):8280000 dirty pages:3189974 colo_bitmap_clear_dirty time consuming(ns):8352000 dirty pages:3189471 colo_bitmap_clear_dirty time consuming(ns):8348000 dirty pages:3189681 colo_bitmap_clear_dirty time consuming(ns):8426000 it can be seen from the data that colo_bitmap_clear_dirty is more efficient. Signed-off-by: Lei Rao <lei.rao@intel.com> Reviewed-by: Lukas Straub <lukasstraub2@web.de> Tested-by: Lukas Straub <lukasstraub2@web.de> --- migration/ram.c | 36 +++++++++++++++++++++++++++++++----- 1 file changed, 31 insertions(+), 5 deletions(-) diff --git a/migration/ram.c b/migration/ram.c index 8661d82..11275cd 100644 --- a/migration/ram.c +++ b/migration/ram.c @@ -857,6 +857,36 @@ unsigned long colo_bitmap_find_dirty(RAMState *rs, RAMBlock *rb, return first; } +/** + * colo_bitmap_clear_dirty:when we flush ram cache to ram, we will use + * continuous memory copy, so we can also clean up the bitmap of +contiguous + * dirty memory. + */ +static inline bool colo_bitmap_clear_dirty(RAMState *rs, + RAMBlock *rb, + unsigned long start, + unsigned long num) { + bool ret; + unsigned long i = 0; + + /* + * Since flush ram cache to ram can only happen on Secondary VM. + * and the clear bitmap always is NULL on destination side. + * Therefore, there is unnecessary to judge whether the + * clear_bitmap needs clear. + */ + QEMU_LOCK_GUARD(&rs->bitmap_mutex); + for (i = 0; i < num; i++) { + ret = test_and_clear_bit(start + i, rb->bmap); + if (ret) { + rs->migration_dirty_pages--; + } + } + + return ret; +} + static inline bool migration_bitmap_clear_dirty(RAMState *rs, RAMBlock *rb, unsigned long page) @@ -3774,11 +3804,7 @@ void colo_flush_ram_cache(void) num = 0; block = QLIST_NEXT_RCU(block, next); } else { - unsigned long i = 0; - - for (i = 0; i < num; i++) { - migration_bitmap_clear_dirty(ram_state, block, offset + i); - } + colo_bitmap_clear_dirty(ram_state, block, offset, num); dst_host = block->host + (((ram_addr_t)offset) << TARGET_PAGE_BITS); src_host = block->colo_cache -- 1.8.3.1
Hi, Dave I think this set of patches is beneficial to upstream. Please check these performance data. If you have any other ideas, please let me know. Thanks Lei. -----Original Message----- From: Rao, Lei Sent: Friday, April 16, 2021 3:57 PM To: dgilbert@redhat.com Cc: qemu-devel@nongnu.org; Zhang, Chen <chen.zhang@intel.com>; lizhijian@cn.fujitsu.com; jasowang@redhat.com; quintela@redhat.com; pbonzini@redhat.com; lukasstraub2@web.de Subject: RE: [PATCH v6 09/10] Add the function of colo_bitmap_clear_dirty Hi, Dave The performance data has added to the commit messages. Do you have any other suggestions? Thanks Lei. -----Original Message----- From: Rao, Lei <lei.rao@intel.com> Sent: Friday, April 9, 2021 11:21 AM To: Zhang, Chen <chen.zhang@intel.com>; lizhijian@cn.fujitsu.com; jasowang@redhat.com; quintela@redhat.com; dgilbert@redhat.com; pbonzini@redhat.com; lukasstraub2@web.de Cc: qemu-devel@nongnu.org; Rao, Lei <lei.rao@intel.com> Subject: [PATCH v6 09/10] Add the function of colo_bitmap_clear_dirty From: "Rao, Lei" <lei.rao@intel.com> When we use continuous dirty memory copy for flushing ram cache on secondary VM, we can also clean up the bitmap of contiguous dirty page memory. This also can reduce the VM stop time during checkpoint. The performance test for COLO as follow: Server configuraton: CPU :Intel(R) Xeon(R) Gold 6140 CPU @ 2.30GHz MEM :251G(type:DDR4 Speed:2666 MT/s) SSD :Intel 730 and DC S35x0/3610/3700 Series SSDs dirty pages:3189376 migration_bitmap_clear_dirty time consuming(ns):105194000 dirty pages:3189784 migration_bitmap_clear_dirty time consuming(ns):105297000 dirty pages:3190501 migration_bitmap_clear_dirty time consuming(ns):105410000 dirty pages:3188734 migration_bitmap_clear_dirty time consuming(ns):105138000 dirty pages:3189464 migration_bitmap_clear_dirty time consuming(ns):111736000 dirty pages:3188558 migration_bitmap_clear_dirty time consuming(ns):105079000 dirty pages:3239489 migration_bitmap_clear_dirty time consuming(ns):106761000 dirty pages:3190240 colo_bitmap_clear_dirty time consuming(ns):8369000 dirty pages:3189293 colo_bitmap_clear_dirty time consuming(ns):8388000 dirty pages:3189171 colo_bitmap_clear_dirty time consuming(ns):8641000 dirty pages:3189099 colo_bitmap_clear_dirty time consuming(ns):8280000 dirty pages:3189974 colo_bitmap_clear_dirty time consuming(ns):8352000 dirty pages:3189471 colo_bitmap_clear_dirty time consuming(ns):8348000 dirty pages:3189681 colo_bitmap_clear_dirty time consuming(ns):8426000 it can be seen from the data that colo_bitmap_clear_dirty is more efficient. Signed-off-by: Lei Rao <lei.rao@intel.com> Reviewed-by: Lukas Straub <lukasstraub2@web.de> Tested-by: Lukas Straub <lukasstraub2@web.de> --- migration/ram.c | 36 +++++++++++++++++++++++++++++++----- 1 file changed, 31 insertions(+), 5 deletions(-) diff --git a/migration/ram.c b/migration/ram.c index 8661d82..11275cd 100644 --- a/migration/ram.c +++ b/migration/ram.c @@ -857,6 +857,36 @@ unsigned long colo_bitmap_find_dirty(RAMState *rs, RAMBlock *rb, return first; } +/** + * colo_bitmap_clear_dirty:when we flush ram cache to ram, we will use + * continuous memory copy, so we can also clean up the bitmap of +contiguous + * dirty memory. + */ +static inline bool colo_bitmap_clear_dirty(RAMState *rs, + RAMBlock *rb, + unsigned long start, + unsigned long num) { + bool ret; + unsigned long i = 0; + + /* + * Since flush ram cache to ram can only happen on Secondary VM. + * and the clear bitmap always is NULL on destination side. + * Therefore, there is unnecessary to judge whether the + * clear_bitmap needs clear. + */ + QEMU_LOCK_GUARD(&rs->bitmap_mutex); + for (i = 0; i < num; i++) { + ret = test_and_clear_bit(start + i, rb->bmap); + if (ret) { + rs->migration_dirty_pages--; + } + } + + return ret; +} + static inline bool migration_bitmap_clear_dirty(RAMState *rs, RAMBlock *rb, unsigned long page) @@ -3774,11 +3804,7 @@ void colo_flush_ram_cache(void) num = 0; block = QLIST_NEXT_RCU(block, next); } else { - unsigned long i = 0; - - for (i = 0; i < num; i++) { - migration_bitmap_clear_dirty(ram_state, block, offset + i); - } + colo_bitmap_clear_dirty(ram_state, block, offset, num); dst_host = block->host + (((ram_addr_t)offset) << TARGET_PAGE_BITS); src_host = block->colo_cache -- 1.8.3.1
diff --git a/migration/ram.c b/migration/ram.c index 8661d82..11275cd 100644 --- a/migration/ram.c +++ b/migration/ram.c @@ -857,6 +857,36 @@ unsigned long colo_bitmap_find_dirty(RAMState *rs, RAMBlock *rb, return first; } +/** + * colo_bitmap_clear_dirty:when we flush ram cache to ram, we will use + * continuous memory copy, so we can also clean up the bitmap of contiguous + * dirty memory. + */ +static inline bool colo_bitmap_clear_dirty(RAMState *rs, + RAMBlock *rb, + unsigned long start, + unsigned long num) +{ + bool ret; + unsigned long i = 0; + + /* + * Since flush ram cache to ram can only happen on Secondary VM. + * and the clear bitmap always is NULL on destination side. + * Therefore, there is unnecessary to judge whether the + * clear_bitmap needs clear. + */ + QEMU_LOCK_GUARD(&rs->bitmap_mutex); + for (i = 0; i < num; i++) { + ret = test_and_clear_bit(start + i, rb->bmap); + if (ret) { + rs->migration_dirty_pages--; + } + } + + return ret; +} + static inline bool migration_bitmap_clear_dirty(RAMState *rs, RAMBlock *rb, unsigned long page) @@ -3774,11 +3804,7 @@ void colo_flush_ram_cache(void) num = 0; block = QLIST_NEXT_RCU(block, next); } else { - unsigned long i = 0; - - for (i = 0; i < num; i++) { - migration_bitmap_clear_dirty(ram_state, block, offset + i); - } + colo_bitmap_clear_dirty(ram_state, block, offset, num); dst_host = block->host + (((ram_addr_t)offset) << TARGET_PAGE_BITS); src_host = block->colo_cache