vecsim index performance: hnswlib::L2SqrSIMD4Ext takes 39.35% of on-cpu cycles ( _mm_add_ps and _mm_loadu_ps are related to it ) #28
Description
Sample rdb loadable via vecsim search brach:
s3://benchmarks.redislabs/redisearch/vecsim/ann-benchmarks/glove-100-angular/dump.rdb
Sample query:
627473461.931705 [0 127.0.0.1:41480] "HSET" "ann_14541" "vector" "\x98Q\xec\xbd\x84\x81\xf7>\xb4<o\xbe*\xe3\xbf>m\xc5\xde\xbe\xf7u\b\xbfK\x02\x14>V\x82%\xbe\x02\x829>W\xb1x<y\x1e$?\xd69\xd6>\xa8Wz?;\xdf\x87\xbf\xa3\x92\xca?=\xb8\xdb\xbe\xcc\xb45?\xdc\x7f\xa4=\x0eJ(?\x97sy?W\xcf\x01?+\xa4\x9c\xbeKY\xa6\xbe\xdb3\xbb>\xee_\xb9>\x9c\xdc\xdf>\xb0\xac4\xbd\x92\a\x82\xbd\xd6s\xb2\xbeU\xde\x1e?E\rN?e\x01\x93<\x1f\xbf7>\xa1\xd6\x14\xbf\x97\xa8\xde\xbe\x1d8\xf7\xbeX9\x84>\xa0\x1a\x97?\x12\xa2\xfc\xbc\xd9|\xac>\x9c\xa7\xba\xbe\xbb\xf2Y>=~\xcf\xbe\x93\xa9\x02>\xa0\xe0\x82\xbe\xffx\xef=u\xc85\xbf\x05\xa2\xa7\xbd\xb1\x16_\xbeJ\x98\t?\xe6t\x19\xbe\xef\xac\xdd=\xf4\xa6:?\xf6E\x02\xbe\x9f\xc8\x0b?ms\x83\xbe\xc8\xef\x05\xbfi5\x04\xbe\xe2X\a?O\xcc\x9a>*o\x17?_{.?8g\x94?\xee\xb1\x84\xbe\xdf\xa6W\xbf>\"\xe6\xbdk\x9a7\xbd\xfb\\M>\x1f\x11\xd3\xbe\\ y?\xcb\xdb1>\"\xc3\x12?+\x18\x95\xbe\xf2\xb5\x1f\xbf\xc4%\xb7\xbe\xb5\x1a\x1a?WC\xa2>\xb9\x88o>$\xd6J?\x89_1\xbc_{F>\x89\xb4\x8d\xbdI.\x8f\xbf\xa2b\x9c>\xaa\xb7\xd6\xbe\xbb\x0f ?z\xc2r>\xa1\x10\xc1\xbe\xda \x03\xbf\xbe\xf6\x8c>\xa1\x83\xae\xbc\xbaI\xb4?\x96&\x85\xbd\xc4\xeb\xca\xbe\xee\xce\x1a?\xf47\x81\xbe\x89\b\xbf\xbd\x99\r\xe2>;p\x8e\xbe/\xf8t="
Top on CPU consumers:
| Flat | Flat% | Sum% | Cum | Cum% | Name | Inlined? |
|---|---|---|---|---|---|---|
| 38718330030 | 17.65% | 17.65% | 38766179198 | 17.67% | _mm_loadu_ps | (inline) |
| 18647707936 | 8.50% | 26.15% | 18666112242 | 8.51% | _mm_add_ps | (inline) |
| 15234704835 | 6.94% | 33.09% | 86321125827 | 39.35% | hnswlib::L2SqrSIMD4Ext | |
| 10545240828 | 4.81% | 37.90% | 10556520348 | 4.81% | _mm_mul_ps | (inline) |
| 3086731940 | 1.41% | 39.31% | 3086731940 | 1.41% | _mm_sub_ps | (inline) |
| 0 | 0.00% | 39.31% | 86317305756 | 39.34% | vectorIndexer | (inline) |
| 0 | 0.00% | 39.31% | 86317305756 | 39.34% | moduleNotifyKeyspaceEvent | |
| 0 | 0.00% | 39.31% | 86317305756 | 39.34% | indexBulkFields | |
| 0 | 0.00% | 39.31% | 68849073985 | 31.38% | hnswlib::HierarchicalNSW::searchBaseLayer | |
| 0 | 0.00% | 39.31% | 12029285653 | 5.48% | hnswlib::HierarchicalNSW::mutuallyConnectNewElement | |
| 0 | 0.00% | 39.31% | 86258259186 | 39.32% | hnswlib::HierarchicalNSW::addPoint | |
| 0 | 0.00% | 39.31% | 86317305756 | 39.34% | Indexes_UpdateMatchingWithSchemaRules | |
| 0 | 0.00% | 39.31% | 86317305756 | 39.34% | Indexer_Process | |
| 0 | 0.00% | 39.31% | 86317305756 | 39.34% | Indexer_Add | |
| 0 | 0.00% | 39.31% | 86317305756 | 39.34% | IndexerBulkAdd | |
| 0 | 0.00% | 39.31% | 86317305756 | 39.34% | IndexSpec_UpdateDoc | |
| 0 | 0.00% | 39.31% | 86317305756 | 39.34% | HashNotificationCallback | |
| 0 | 0.00% | 39.31% | 86317305756 | 39.34% | HNSWIndex_AddVector | |
| 0 | 0.00% | 39.31% | 86317305756 | 39.34% | Document_AddToIndexes | |
| 0 | 0.00% | 39.31% | 86317305756 | 39.34% | AddDocumentCtx_Submit |
Flame Chart detail of hnswlib::L2SqrSIMD4Ext cpu cycles
