On 7/30/2011 2:28 AM, Bryan Duxbury wrote:
> If you want efficient data structure for binary data, use the "binary"
> thrift data type. It maps to ByteBuffer backed by a byte[] in Java.
>
I have made some detailed measurement of performance.
typedef list<byte> OctetSeq
service ATL {
i16 octet_thruput(1:OctetSeq payload),
i16 octet_thruput2(1:binary payload),
}
Thus, I have two types of payload and have tested this RPC performance
using 2 languages.
Java (java_client<-->java_server) and
cpp (cpp_client<-->cpp_server)
This gives me 4 results in all. The size of payload I have varied from 4
bytes to 64K bytes. For each call I measure round trip latency. I did 1
million roundtrips for all but the "Java list<byte>" case, there I did
10,000 round trips in order to save time. Client and server are on two
Dell PC1950 Fedora 12 machines on isolated network, GigEthernet. (Java
and C++ TCP results are identical.)
As one can see from attached graphic the "list<byte>" payload in java
is about 100x slower than same thrift RPC being made in C++. Surely it
seems one should find to efficient mappings possible, no? While in place
of "list<byte>" one can use "binary", there is no similar substitute for
"list<i32>" or "list<double>". So it seems that for
"list<primitive_data_type>" mapping should use either
ByteBuffer,DoubleBuffer,Int/Long/ShortBuffer, no?
Gautam
