The full report is here. Copying from the Summary...
This has results for IO-bound Linkbench. By IO-bound I mean the database is much larger than RAM. Tests were run on a single-node replica in DSI. The database should be ~300G after the load. The load uses ~21 threads and transaction tests used 1, 16, 32 and 64 client threads.
Results are provided for 4.0, 4.2 and two 4.4 builds. For 4.4 there is 4.4pre which is a recent 4.4 build prior to the durable history merge using this URL and then 4.4dh using this URL and git hash 00502b which is the first commit with durable history.
Comparing 4.4dh with 4.4pre
- 17% regression in the load rate for 4.4dh
- ~20% regression in the transaction/second rates for 4.4dh
- ~2X regression in mongod RSS for 4.4dh, this looks like a memory leak
- 31% regression in the database size for 4.4dh
The memory leak and disk space regressions are visible during the load.
This test used the new (optimized) Linkbench client, called linkbench2 in DSI. BF-16496 is open for a regression in classic Linkbench, although that was run for a shorter time and with a smaller database and the memory leak and database size problems are easier to spot here.