SO_TIMESTAMP may go away

[Hal Murray]

dfoxfranke at gmail.com said:
> If you try to measure the cost of the authentication code using log messages
> you're going to get total noise, because the cost of logging a message is
> higher than the cost of doing the authentication. Each invocation of AES-SIV
> should take, in round numbers, 250 CPU cycles, and processing a typical
> request should take four invocations (unseal cookie, unseal message, seal new
> cookie, seal message), with only the final one being in the critical section.

I was planning to collect a few time stamps while processing a message, then 
write out the log message after sending the packet.  Under light loads, the 
extra cost of a log message shouldn't be a problem.

But that won't work for interesting loads.  I'm scratching my head to come up 
with a way to preprocess a batch before writing things out.

Is generating randomness likely to be significant?


-- 
These are my opinions.  I hate spam.