You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Will probably need to test on low dimension vectors (i.e. random data, not the grid search) so that vector comparison time doesn't swamp the Map in the measurement.
The text was updated successfully, but these errors were encountered:
@jbellis I did some synthetic testing in the cases I'd expect to stress the difference. Workloads included concurrency levels of 4/8/16/32/64 threads with mixed read-write workloads profiles consisting of 12.5% writes, 25% writes, and 50% writes. In addition, I also tested concurrency levels of 8/16/32 prebuilding the graph (all writes) and searching random vectors (all reads).
Profiles confirmed this set up did cause data structure operations to dominate search/indexing, as desired. In all cases, the differences between CHM and NBHM were minimal and within run-to-run variation. Given this, I think there's no reason to advocate for a change at this time. We could revisit this down the road, but I think there are plenty of higher payoff changes right now.
See commit history for where NBHM got added.
Will probably need to test on low dimension vectors (i.e. random data, not the grid search) so that vector comparison time doesn't swamp the Map in the measurement.
The text was updated successfully, but these errors were encountered: