When __wt_log_flush waits for earlier closed log slots to finish writing, it sleeps for at least 1 ms and forces out the active slot on every iteration. This change replaces the sleep and redundant forced slot switches with a timed wait of up to 200 usec on log->log_write_cond followed by __wti_log_wrlsn. This removes unnecessary slot writes and cuts the wait loop duration from ~1.5 ms to ~0.31 ms under slow sync operations.

Problem

Line numbers reference develop 50c032ba6. __wt_log_flush (src/log/log.c:2840-2881) is called by WT_SESSION::log_flush (src/session/session_api.c:1213) and by checkpoints when logging is enabled (src/checkpoint/checkpoint_txn.c:1560). It forces out the active slot, assigns alloc_lsn to last_lsn (log_close_lsn if the log moved to a new file), and waits until write_lsn reaches it (log.c:2866-2869):

while (__wt_log_cmp(&last_lsn, &lsn) > 0) {
    __wt_sleep(0, WT_THOUSAND);
    WT_RET(__wt_log_flush_lsn(session, &lsn, false));
}

The loop runs when a slot below last_lsn has closed but has not finished writing. Worker threads close slots when a write reaches half the slot buffer size (log.c:2721-2723, 128 KB with the default 256 KB buffer, log_slot.c:466-467) or for records exceeding that threshold (log_slot.c:574-576). The last thread to release the slot writes it out.

Each loop iteration currently incurs:

• A sleep of at least 1 ms (select() with a 1000 usec timeout took 1095 usec at p50 on the Linux host and 1511 usec on the loaded Mac host).
• A redundant forced switch of the active slot: acquiring log_slot_lock, looping on EBUSY, and issuing a pwrite of that slot under the lock (log_slot.c:376-383).
• A call to __wti_log_wrlsn under log_writelsn_lock.

The additional forced slot switch is unnecessary. At loop entry, all slots below last_lsn are already closed (alloc_lsn only advances when a slot closes, log_slot.c:187), and the thread releasing each slot performs the write. Slots released without WTI_SLOT_FLUSH or WTI_SLOT_SYNC* are marked written without signaling (log.c:1949-1958). write_lsn advances when the wrlsn thread or a log_flush caller executes __wti_log_wrlsn (log_mgr.c:754-872), which signals log_write_cond on each advance (log_mgr.c:860).

The same loop exists in mongodb-8.0 (cd0025dd3, src/log/log.c:2854).

When it shows up in MongoDB

JournalFlusher is the only thread that makes the journal durable. Under steady w:"majority" or j:true write workloads, waitUntilDurable calls WT_SESSION::log_flush("sync=on") back-to-back, with callers waiting for the ongoing round plus the subsequent round. Because mongod does not set log.file_max, the slot buffer defaults to 256 KB. The loop adds at least 1 ms per round whenever more than 128 KB of journal data arrives during the round, or when an individual oplog entry exceeds 128 KB.

Measurements

Microbenchmark running directly against WiredTiger. N writer threads each commit one 200-byte insert per transaction (one 256-byte log record, transaction_sync disabled, log=(file_max=100MB)). A flusher thread calls log_flush("sync=on") as soon as at least one transaction commits since the previous call started, modeling JournalFlusher behavior. Runs used a 2 s warm-up and 8 s measurement period, alternating build order across rounds. Evaluated on release builds of 50c032ba6 (develop) and develop with the patch (branch). Cells report medians across rounds, with [min-max] ranges where noted.

Environment:

Mac, 5 rounds:

Mac, wait loop isolated with temporary instrumentation counters, 3 rounds. "signal" includes this change plus signaling wrlsn.cond prior to each wait, matching __log_wait_for_earlier_slot:

Signaling wrlsn.cond does not decrease loop time. It wakes the wrlsn thread, which then yields up to 1000 times without work (log_write_lsn_skip is 8-14x higher than branch). The patch therefore does not signal wrlsn.cond.

dev-m (Linux), series A, 5 rounds. "Loop iterations per call" for develop is calculated as (log_force_write - log_flush) / log_flush minus the corresponding branch value:

dev-m, series A2, 8 rounds, measuring develop twice to capture run-to-run variation:

dev-m, series L: 4 writers, 64 KB records, 3,000 commits/s total (188 MB/s log output), 6 rounds:

On the Mac, the loop executes in 82-91% of log_flush calls, taking 14-17% of total flush time. The change lowers loop time from ~1.5 ms to ~0.31 ms and improves log_flush p50 by 10-35%. On dev-m, the loop executes in only 1-6% of calls (~10% with 64 KB records) because fdatasync is fast enough that closed slots are written before the next round begins; performance deltas stay within develop run-to-run variance (up to 12% at p50). The change benefits workloads where slot writes are slow relative to a flush round, such as on APFS with F_FULLFSYNC, with no measurable overhead when the loop is bypassed.

Proposed change

PR: https://github.com/wiredtiger/wiredtiger/pull/14950.
In the wait loop, wait on log->log_write_cond for up to 200 usec (the same timeout used by __log_wait_for_earlier_slot), call __wti_log_wrlsn, and re-read write_lsn. Check for panics before waiting. Do not force out the active slot and do not signal the wrlsn thread. The loop exit condition, log_close_lsn handling, and __wt_log_force_sync remain unchanged.

Local testing

Tested on the PR branch, based on 50c032ba6. macOS arm64, Apple clang 21.0.0, RelWithDebInfo with HAVE_DIAGNOSTIC=1, ENABLE_STRICT=1 (-Werror), and Python:

• run.py -j 4 txn log test_backup07.py test_backup09.py test_backup10.py test_backup24.py test_backup25.py test_prepare02.py test_readonly03.py test_schema08.py test_timestamp03.py test_timestamp07.py test_truncate21.py: Ran 458 tests, OK (skipped=13).
• ctest ex_log, test_truncated_log, test_random_abort, test_timestamp_abort, test_schema_abort: 5 of 5 passed.
• test/format/t -1 logging=1 logging.remove=1 runs.timer=2 runs.type=row checkpoint=on: passed (29 tables, 20 threads, logging.file_max=499635).
• dist/s_all with clang-format 23.1.2: no modifications to src/log/log.c. Unrelated reformats in bench/workgen/workgen.cpp, ext/page_log/palite/palite.cpp, and test/catch2/misc_tests/test_rec_upd_select.cpp were discarded. ruff_check.py and s_docs were skipped as Ruff and Doxygen are not installed; this change touches no Python or documentation files.