lmb
commented
8 years ago So, here is how I interpret this:
- The for loop runs N times
- For N runs, makeBenchDBVal makes 21 allocations
- For N runs, Txn.Put makes 440404 allocations
Now, considering that N is kind of large (a few 100k probably), allocations from makeBenchDBVal do not factor in. The 4 allocations per op are exclusively from Txn.Put. Not very exciting so far.
But where are those allocations happening in Txn.Put (this is a different run from the one above, so absolute numbers won't match up)?
Total: 906361
ROUTINE ======================== github.com/bmatsuo/lmdb-go/lmdb.(*Txn).Put in /Users/lorenz/go/src/github.com/bmatsuo/lmdb-go/lmdb/txn.go
905940 905940 (flat, cum) 100% of Total
. . 314: vn := len(val)
. . 315: if vn == 0 {
. . 316: val = []byte{0}
. . 317: }
. . 318:
905940 905940 319: ret := C.lmdbgo_mdb_put2(
. . 320: txn._txn, C.MDB_dbi(dbi),
. . 321: unsafe.Pointer(&key[0]), C.size_t(kn),
. . 322: unsafe.Pointer(&val[0]), C.size_t(vn),
. . 323: C.uint(flags),
. . 324: )- L320: No allocations here? Copied to the stack
- L321: 1 allocation?
- L322: 1 allocation?
- L323: No allocations again
The only way I can make the numbers match up is if unsafe.Pointer(&key[0]) does two allocations. Which I don't understand.
bmatsuo
@lmb pointed out some anomalous looking numbers reported by
go test -bench=. -benchmemin #61, Specifically regarding the BenchmarkTxn_Put function.The 4 allocations reported are difficult to track down.
One should be happening in the benchmark function itself(edit: that is false, allocations in the benchmark function get amortized and do not have a significant effect). But the others are not easy to account for. In a normal (non-error) case Txn.Put does not appear to make allocations inside the Go runtime.@lmb ran the pprof tool using
-alloc_objectsand the following information about allocations in the benchmark function was comptued:Nothing seems to be adding up correctly here.