Some background: while switching to write_buf in hyper, I noticed the 16 pipelined requests benchmark suffered since the responses where split of 32 buffers (1 headers + 1 body * 16 responses), and so each syscall was only writing half as many responses as when hyper flattens all into a single buffer first.
The easiest knee-jerk reaction is to increase the amount of IoVecs used in TcpStream::write_buf. I arbitrarily picked 64. The man page for writev(2) mentions that the default max in Linux is 1024...
A longer term fix could be to add an associated constant to Buf, to allow an implementer to state how many IoVecs they expect to need.
Some background: while switching to
write_bufin hyper, I noticed the 16 pipelined requests benchmark suffered since the responses where split of 32 buffers (1 headers + 1 body * 16 responses), and so each syscall was only writing half as many responses as when hyper flattens all into a single buffer first.The easiest knee-jerk reaction is to increase the amount of
IoVecs used inTcpStream::write_buf. I arbitrarily picked 64. The man page forwritev(2)mentions that the default max in Linux is 1024...A longer term fix could be to add an associated constant to
Buf, to allow an implementer to state how manyIoVecs they expect to need.