Most of the time it's most efficient to batch work.

But.. if you have multiple threads this can be difficult to coordinate, especially if the threads need to respond to clients quickly.

bmon (aka batching monitor) makes it easy to batch work from multiple threads.

Case study: Batching PearDB's LMDB writes

LMDB only allows a single writer. To get around this it's best to batch your writes into a single transactionn.

This is what PearDB does.

Checkout the sweet speed up:

Before

bench [master] λ wrk -d 10 -c 30 -t 30 --latency -s writeonly-pipelining.lua http://127.0.0.1:8888                                                                                              ~/d/p/bench
Running 10s test @ http://127.0.0.1:8888
  30 threads and 30 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency   154.77ms    2.34ms 173.90ms   98.00%
    Req/Sec   102.79      1.80   107.00     79.00%
  Latency Distribution
     50%  154.57ms
     75%  155.03ms
     90%  155.81ms
     99%  173.10ms
  30947 requests in 9.98s, 5.11MB read
Requests/sec:   3102.23
Transfer/sec:    524.11KB

After

bench [feature/batch_write-rebase] λ wrk -d 10 -c 30 -t 30 --latency -s writeonly-pipelining.lua http://127.0.0.1:8888                                                                           ~/d/p/bench
Running 10s test @ http://127.0.0.1:8888
  30 threads and 30 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency    17.61ms    4.06ms  36.44ms   92.52%
    Req/Sec     0.94k   130.62     1.18k    90.95%
  Latency Distribution
     50%   16.59ms
     75%   17.27ms
     90%   18.22ms
     99%   33.10ms
  269407 requests in 9.97s, 44.45MB read
Requests/sec:  27009.81
Transfer/sec:      4.46MB