Use bounce buffer ring to optimize local pread

[kingcrimsontianyu]

Related PRs and issues

Depends on #913 #919
Addresses #914

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[kingcrimsontianyu]

/ok to test

[kingcrimsontianyu]

/ok to test

[kingcrimsontianyu]

Performance check

Sequential read

• ipp1-3304

  • main: syscall pread and host-to-device copy do not overlap

• This PR, ring with 4 buffers: syscall pread and host-to-device copy overlap

• Hot BW: read bandwidth when the data is in the page cache
• Cold BW: read bandwidth when the data has been cleared from the page cache
• Cold BW: same with the above except that KVIKIO_AUTO_DIRECT_IO_READ=1 which makes opportunistic use of Direct I/O.

Branch	Threads	Hot BW [GiB/s]	Cold BW [GiB/s]	Cold BW (auto direct I/O) [GiB/s]
main	1	3770.7933	1993.8511	4027.8119
main	4	13368.3568	4225.3567	6450.4380
ring (1 buffer)	1	4086.5414	1885.2542	3890.0020
ring (1 buffer)	4	13336.1997	4280.2705	6463.3447
ring (4 buffers)	1	4165.6537	1852.0735	4019.2380
ring (4 buffers)	4	14927.6879	4745.0288	6466.7356

[kingcrimsontianyu]

/ok to test