In Java + Linux, what is the correct way to measure latency across CPU cores? -

i've heard system.nanotime() not consistent across cpu cores in java + linux. how can measure latency of one-way message on loopback, in other words, in same machine. flow is:

• client sends message server timestamp in it
• server gets message
• servers parses timestamp out of message
• servers calculates one-way latency (now - timestamp)

note server , client running in same machine pinned different cpu cores.

you can use system.nanotime(). it's problem on old oses such multi-socket machines running windows xp.

within machine.

• write timestamp in 1 thread
• read timestamp in thread.

note: measuring latency way has serious problems such co-ordinated omission. better approach is

• calculate when test should start based on target throughput want measure.
• busy wait time reached if not behind.
• write time message should have been sent rather time sent.
• read timestamp , compare.
• record distribution.

you can use system.nanotime() between machines doing round trip test

• send timestamp
• have other end send back.
• compare current time.