|
Description:
The patch improves the performance of vectored raw I/O by combining
the iovecs passed by a readv/writev call into a single kiobuf and
issuing one call to brw_kiovec to submit the io. Without the patch,
the readv/writev system calls (effectively) serially invoke the
read/write system calls for each iovec.
With the patch applied, raw readv/writev approximates the speed
of a read/write for the aggregate iov_len's (rather than vectored
I/O running 2 or 3 times slower). In preliminary DB2 testing, the
time for a single tablescan was reduced from 621 seconds to 172
seconds with the patch applied.
|