Evan Jones wrote:
Problem 2: Using the NIO encoders/decoders can be faster than
String.getBytes, but only if it is used >= 4 times. If used only
once, it is worse. The same is approximately true about decoding.
Lame results: http://evanjones.ca/software/java-string-encoding.html
I'm revisiting this old issue, thanks to being reminded about it by an
earlier message. I've tested this with "recent" JVMs, and it still
seems to hold true: using the NIO encoders and decoders can be faster
than using String.getBytes(). My numbers show that encoding is
approximately 40% faster, while decoding shows a smaller improvement.
See the following for details on this microbenchmark:
http://evanjones.ca/software/java-string-encoding.html
This surprises me, since it suggests that Sun/Oracle could replace
their implementation of String.getBytes() with something similar to my
code, and get a performance improvement. In fact, with privileged
access to the internals of a String, they should be able to do even
better.
I've integrated this change into protobuf and with the microbenchmark
in the protobuf source tree, it shows a performance improvement of
~20% (numbers below). I'll send a code review with the code shortly,
once I've cleaned it up a bit, in case anyone wants to look at it.
Pros:
+ Faster protocol buffer encoding.
Cons:
- Far more code to handle encoding (like an extra hundred lines or so)
which means there could be bugs.
- Extra memory (~2 kB per thread for encoding)
I'm unsure if the benefits outweigh the costs, particularly since the
fact this appears faster seems to be a surprising result, and I
wouldn't be shocked to find that future JVM / JDK releases could make
this "optimization" useless. I'll leave that decision to the protocol
buffer maintainers.
Evan
Results: All the serialize results are better. The deserialized
results are unchanged (as expected).
SpeedMessage1 (which is small):
Original: Serialize to byte string: 12360424 iterations in 29.83s;
90.097984MB/s
Optimized: Serialize to byte string: 15911623 iterations in 29.997s;
115.337776MB/s
SpeedMessage2 (larger):
Serialize to byte string: 33482 iterations in 29.754s; 90.757484MB/s
Serialize to byte string: 40381 iterations in 30.031s; 108.44853MB/s
Raw results on my Macbook (Core2 Duo CPU):
ORIGINAL
Benchmarking benchmarks.GoogleSpeed$SpeedMessage1 with file
google_message1.dat
Serialize to byte string: 12360424 iterations in 29.83s; 90.097984MB/s
Serialize to byte array: 12244951 iterations in 30.303s; 87.86307MB/s
Serialize to memory stream: 8699469 iterations in 23.732s; 79.70642MB/s
Serialize to /dev/null with FileOutputStream: 6075179 iterations in
27.764s; 47.578636MB/s
Serialize to /dev/null reusing FileOutputStream: 6975006 iterations in
29.99s; 50.571175MB/s
Serialize to /dev/null with FileChannel: 10375092 iterations in
29.864s; 75.54034MB/s
Serialize to /dev/null reusing FileChannel: 11166943 iterations in
30.699s; 79.09427MB/s
Deserialize from byte string: 14463117 iterations in 30.06s;
104.618355MB/s
Deserialize from byte array: 14436567 iterations in 30.007s;
104.61074MB/s
Deserialize from memory stream: 6221772 iterations in 28.024s;
48.274624MB/s
Benchmarking benchmarks.GoogleSpeed$SpeedMessage2 with file
google_message2.dat
Serialize to byte string: 33482 iterations in 29.754s; 90.757484MB/s
Serialize to byte array: 33103 iterations in 29.517s; 90.45062MB/s
Serialize to memory stream: 28872 iterations in 29.939s; 77.77786MB/s
Serialize to /dev/null with FileOutputStream: 32934 iterations in
29.927s; 88.756MB/s
Serialize to /dev/null reusing FileOutputStream: 32979 iterations in
29.887s; 88.99622MB/s
Serialize to /dev/null with FileChannel: 32447 iterations in 29.921s;
87.46108MB/s
Serialize to /dev/null reusing FileChannel: 32585 iterations in
29.903s; 87.88594MB/s
Deserialize from byte string: 38388 iterations in 29.879s;
103.620544MB/s
Deserialize from byte array: 38677 iterations in 29.866s; 104.446075MB/s
Deserialize from memory stream: 37879 iterations in 29.954s;
101.990585MB/s
OPTIMIZED
Benchmarking benchmarks.GoogleSpeed$SpeedMessage1 with file
google_message1.dat
Serialize to byte string: 15911623 iterations in 29.997s; 115.337776MB/s
Serialize to byte array: 16152646 iterations in 30.008s; 117.041954MB/s
Serialize to memory stream: 14859367 iterations in 29.551s;
109.33597MB/s
Serialize to /dev/null with FileOutputStream: 7224915 iterations in
29.954s; 52.446056MB/s
Serialize to /dev/null reusing FileOutputStream: 7479144 iterations in
30.081s; 54.062305MB/s
Serialize to /dev/null with FileChannel: 12730586 iterations in
30.025s; 92.193504MB/s
Serialize to /dev/null reusing FileChannel: 14024645 iterations in
30.399s; 100.31538MB/s
Deserialize from byte string: 14390338 iterations in 29.958s;
104.44631MB/s
Deserialize from byte array: 14496442 iterations in 30.142s;
104.57414MB/s
Deserialize from memory stream: 6653401 iterations in 29.989s;
48.24104MB/s
Benchmarking benchmarks.GoogleSpeed$SpeedMessage2 with file
google_message2.dat
Serialize to byte string: 40381 iterations in 30.031s; 108.44853MB/s
Serialize to byte array: 39447 iterations in 29.678s; 107.20024MB/s
Serialize to memory stream: 33647 iterations in 30.029s; 90.36951MB/s
Serialize to /dev/null with FileOutputStream: 39071 iterations in
30.142s; 104.543945MB/s
Serialize to /dev/null reusing FileOutputStream: 39321 iterations in
29.969s; 105.82024MB/s
Serialize to /dev/null with FileChannel: 38460 iterations in 29.955s;
103.55149MB/s
Serialize to /dev/null reusing FileChannel: 38690 iterations in
29.969s; 104.12209MB/s
Deserialize from byte string: 37440 iterations in 29.312s;
103.016495MB/s
Deserialize from byte array: 38208 iterations in 29.862s; 103.193375MB/s
Deserialize from memory stream: 37315 iterations in 29.944s;
100.505554MB/s
--
Evan Jones
http://evanjones.ca/
--
You received this message because you are subscribed to the Google Groups "Protocol
Buffers" group.
To post to this group, send email to [email protected].
To unsubscribe from this group, send email to
[email protected].
For more options, visit this group at
http://groups.google.com/group/protobuf?hl=en.
