Den mån 3 maj 2021 23:43Tim Newsome <t...@sifive.com> skrev: > Change > https://sourceforge.net/p/openocd/code/ci/7dd323b26d93e49e409e02053e30f53ac8138cd5/ > cut remote bitbang performance (when talking to the spike RISC-V simulator) > approximately in half. This change "removes the file write stream, replaces > it with socket write calls." Previously performance was better because the > individual byte-size writes were buffered, and didn't result in a system > call until fflush() was called. > > I assume the right way to fix this is to implement a buffer/flush > mechanism inside OpenOCD, on top of the socket calls, so that the > optimization also works on Windows. Is that right? Is anybody motivated to > take that on? > > Tim >
Hi! It feels entirely unnecessary to implement a buffering mechanism if the C library already has a perfectly good implementation. I think either live with the performance degradation*, or reimplement the mentioned change to use socket calls only on windows, via a common wrapper and some autotools magic. * the TCP socket should be buffering small writes as well, via the Nagle algorithm. Is perhaps TCP_NODELAY set on the socket? /Andreas >
