The three awk scripts in a trench coat have pretended to be a grown-up tool for long enough to point the way, though the analysis portion is definitely doing the "is this version's measurement better or worse than the previous version" calculation the wrong way so I'm kind of just eyeballing the results.
Among other things, the next version of Pipe Viewer will extend the "%T" / "%{buffer-percent}" format string to indicate the transfer method - read/write, splice, splice via intermediate pipe, or copy-file-range. That's what eventually showed the cause of the performance drop.
It turns out that if you pass a length of SIZE_MAX to splice() or copy_file_range(), then sometimes the wrong things happen and the transfer has to fall back to read()/write().
Now if only I could work out why on CentOS 5 - and only there - the "integrity when adjusted remotely" test has started hanging until you attach "strace" to the running "pv" process. Maybe that's a butterfly to chase some other time.
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.