Project

General

Profile

Bug #38538

Updated by Jason Dillaman about 1 year ago

The "object_map_update" OSD class method has a some low-hanging improvements that can be made to increase the performance of object-map updates:

1) the std::vector BitVector used uses for the bitset data is zeroed before it's contents are replaced with the bitset data read from the object-map object. This is a huge unnecessary CPU hit for larger object-maps.

2) the entire footer is read into memory even though only 4 bytes might be changed for an updated CRC. For large RBD images, this could be up to 16KiB of data in the worst-case scenario.

A stress-test against librados_test_stub shows improvement from ~100 updates/sec to over 50K updates/sec. The OSD is still limited to approximately 2-3K write operations per second, so longer term refactoring of the object-map design will be required to improve it's performance by orders of magnitude.

Back