binderRpcBenchmark: device tests kernel as ctrl
Host still tests RPC binder only since kernel binder is only on Android
(by convention!).
On local cuttlefish:
Tests suffixes:
\0 is KERNEL
\1 is RPC
...
---------------------------------------------------------------
Benchmark Time CPU Iterations
---------------------------------------------------------------
BM_pingTransaction/0 28448 ns 18366 ns 36940
BM_pingTransaction/1 39291 ns 26773 ns 25295
BM_repeatString/0 184458 ns 96268 ns 7263
BM_repeatString/1 206385 ns 116369 ns 5947
BM_repeatBinder/0 50475 ns 43567 ns 15896
BM_repeatBinder/1 50722 ns 44386 ns 15857
So the takeaway is that RPC binder has some overhead for very small
transactions.
Future considerations:
- make the RPC calls across processes instead of across threads
- see how scheduler optimizations affect this
Fixes: 194903439
Test: binderRpcBenchmark
Change-Id: Ib7b680091979d142bca4716d62cc579f90bade0f
1 file changed