Skip to content

Comment on The C10M problem (2013)parent

Comments

Yes, but this puts a lower limit on the latency of each individual request, not on the throughput. Consider having 4 3-level tree lookups that you need to perform. If the first level hits L1, the second L3, and the 3rd is in RAM, you'd have about 10 + 40 + 100 = ~150 cycles of minimum latency for each request.

Serially, you'd have ~600 cycles cycles of total latency for the 4 lookups. But if you can do the lookups in batches of 4, you can overlap your latencies and increase your throughput. You can issue the 4 parallel first level lookups, then 4 second level, and then the 4 third level in very close to the same ~150 cycles that you require to do a single complete lookup.

AboutSource Built by g1lg1l

Hackerly is an independent reader for Hacker News, built on the public HN API. Not affiliated with Y Combinator.