Hi,
To measure throughput, I'm timing exec time spent in LibPQ,
against the size of the result-set in bytes, as reported by PQresultMemorySize().
EXEC: 7x ( 130,867 rows, 54,921,532 bytes) in 0.305s (171.8 MB/s)
EXEC: 8x ( 180,079 rows, 95,876,047 bytes) in 0.493s (185.5 MB/s)
EXEC: 9x ( 224,253 rows, 371,663,836 bytes) in 2.986s (118.7 MB/s)
The problem is that I have only a 1GbE network link, so theoretically,
max-throughput should be around 125MB/s, which the 1st two runs exceed.
These 3 runs access the same schema, doing full scans of a few "real data" tables,
and the last one accesses more/larger bytea columns. These are plain SELECTs in
binary mode using normal sync execution (no cursors, COPY, single-row mode, PIPELINE, etc...)
Obviously from these results, I now realize PQresultMemorySize() returns
something larger than what went across the network. Can someone explain
how so? And if there's a better proxy to programmatically know the network
traffic exchanged on the connection's socket, that's cross-platform? Obviously
libpq itself knows, but I don't see any way to access that info.
Perhaps tracing might? But will that incur overhead?
I'd appreciate any insight. Thanks, --DD