gomemcached
This is a memcached binary protocol toolkit in go.
It provides client and server functionality as well as a little sample server showing how I might make a server if I valued purity over performance.
Server Design
The basic design can be seen in gocache. A storage
server is run as a goroutine that receives a MCRequest on
a channel, and then issues an MCResponse to a channel contained
within the request.
Each connection is a separate goroutine, of course, and is responsible
for all IO for that connection until the connection drops or the
dataServer decides it's stupid and sends a fatal response back over
the channel.
There is currently no work at all in making the thing perform (there are specific areas I know need work). This is just my attempt to learn the language somewhat.
This fork
This fork attempts to avoid extraneous copies when transmitting objects by matching the in-memory representation with the on-wire format. The performance improvements are ~10x for transmits:
benchmark old ns/op new ns/op delta
BenchmarkEncodingRequest 194 19.2 -90.10%
BenchmarkEncodingRequest0CAS 186 19.5 -89.52%
BenchmarkEncodingRequest1Extra 190 19.5 -89.74%
BenchmarkReceiveRequest 212 241 +13.68%
BenchmarkReceiveRequestNoBuf 297 363 +22.22%
BenchmarkEncodingResponse 195 19.1 -90.21%
BenchmarkEncodingResponseLarge 28665 19.2 -99.93%
BenchmarkReceiveResponse 203 242 +19.21%
BenchmarkReceiveResponseNoBuf 296 357 +20.61%
benchmark old MB/s new MB/s speedup
BenchmarkEncodingRequest 205.21 2086.10 10.17x
BenchmarkEncodingRequest0CAS 214.14 2049.70 9.57x
BenchmarkEncodingRequest1Extra 215.02 2100.60 9.77x
BenchmarkReceiveRequest 193.09 169.65 0.88x
BenchmarkReceiveRequestNoBuf 137.96 112.87 0.82x
BenchmarkEncodingResponse 204.76 2092.48 10.22x
BenchmarkEncodingResponseLarge 858.43 1279381.59 1490.37x
BenchmarkReceiveResponse 201.10 168.88 0.84x
BenchmarkReceiveResponseNoBuf 138.07 114.75 0.83x