The following HelloWorld app was used as benchmark:
data HelloWorld = HelloWorld
mkYesod "HelloWorld" [parseRoutes|
/ HomeR GET
|]
instance Yesod HelloWorld where
-- makeSessionBackend = const $ return Nothing
getHomeR = return . RepPlain . toContent $ "Hello World!"
main :: IO ()
main = warp 8080 HelloWorld
The benchmark was tested with httperf under the following
environments:
[vanilla-nosession] Released yesod-core 1.1.1.1, but without sessions.
[vanilla-session] Released yesod-core 1.1.1.1 (with sessions).
[faster-session-1] With patch 065e33a, "Faster, leaner
implementation of putTime/getTime".
[faster-session-2] With this commit.
Performance results:
A) Testing with:
httperf --hog --client=0/1 --server=localhost \
--port=8080 --uri=/ --rate=1000 \
--send-buffer=4096 --recv-buffer=16384 \
--num-conns=100 --wsess=1000,60,1 \
--burst-length=20 --session-cookie
Results:
vanilla-nosession: 19187.7 req/s (0.1 ms/req)
vanilla-session: 2523.3 req/s (0.4 ms/req)
faster-session-1: 2933.5 req/s (0.3 ms/req)
faster-session-2: 2957.5 req/s (0.3 ms/req)
This test benchmarks both saving and loading the session.
Interestingly, this commit provides only a small performance
increase.
B) Testing with:
httperf --hog --client=0/1 --server=localhost \
--port=8080 --uri=/ --rate=1000 \
--send-buffer=4096 --recv-buffer=16384 \
--num-conns=100 --num-calls=1000 \
--burst-length=20
Results:
vanilla-nosession: 43548.7 req/s (0.0 ms/req)
vanilla-session: 3609.4 req/s (0.3 ms/req)
faster-session-1: 3454.9 req/s (0.3 ms/req)
faster-session-2: 3786.9 req/s (0.3 ms/req)
This test benchmarks only saving the session. Strangely,
faster-session-1 was worse than vanilla-session (while
isolated tests show that the new putTime should be +70x
faster). However, there is a non-negligible performance
increase on faster-session-2.
Benchmark on my computer (per call, includes runPut/runGet):
old putTime: 5658 ns +/- 224ns
new putTime: 821 ns +/- 24ns (7x faster)
old getTime: 7228 ns +/- 126ns
new getTime: 99 ns +/- 4ns (73x faster!!)
Besides, the old format used 25 raw bytes (33.3 bytes on the
base64 output), while the new one uses 8 bytes (10.6 bytes on the
base64 output).
