As a result, it did not show up in pstats. Alos, we now use the input name
in worker thread to easily identify where they belong to. As thread names
have very limited length, the thread naming now is
"w<worker-number>/<input-name>".
This commit significantly enhances imtcp by introducing a fully
functional worker thread pool, enabling true multi-threaded
processing for better scalability under high loads. This is
particularly beneficial when using TLS connections.
Notable changes:
- Implemented a complete worker pool for imtcp.
- Introduced the `workerthreads` config parameter for tuning
concurrency.
- Improved epoll efficiency by enabling edge-triggered mode.
- Added starvation handling via `starvationProtection.maxReads`.
- Refactored session accept logic and optimized network object
handling.
- Removed an obsolete network driver layer for event notification.
- Fixed multiple issues related to message timing, EPOLLERR
handling, and tests.
- Improved performance in poll() mode by reducing redundant
allocations.
- Introduced new CI tests for imtcp without epoll.
- Allowed disabling imtcp tests via a new configure switch.
- Added new impstats counters for worker thread pool statistics.
Details:
- The worker pool replaces an outdated experimental
implementation.
- If `workerthreads=1`, no worker pool is created to minimize
context switches.
- Moves worker pool variables inside `tcpsrv` instance to
prevent conflicts.
- Extracts session `accept()` logic into a dedicated function
for clarity.
- Fixes message ordering inconsistencies in multi-threaded
scenarios.
- Properly handles `EPOLLERR` notifications to improve error
resilience.
- Optimizes poll() mode by avoiding unnecessary reallocation
of file descriptors.
- Replaces the old network driver layer for event notification
with a streamlined solution.
- Now uses **conditional compilation** to select the best
method (epoll or poll) at build time.
- This significantly reduces code complexity, improves
maintainability, and boosts performance.
- The previous "thread pool" was a rough experiment that did
not perform significantly better than single-threaded mode.
- The **new implementation** allows multiple worker threads
on platforms with `epoll`.
- On non-epoll systems, an optimized **poll() based
single-threaded approach** is used, which is expected to
perform better than the old "thread pool."
- Adds `pthread_setname_np` only when available to improve
portability.
- Fixes test cases that assumed strict message timing, which
was unreliable.
- Reduces test parallelism for TSAN CI runs to prevent
resource exhaustion.
- Moves a test case to `imdiag` to ensure stable execution.
- Provides a new CI environment to verify `imtcp` behavior
without epoll.
- Introduces `--enable-imtcp-tests` configure switch for test
flexibility.
- Improves debug logging and adds better error handling for
worker pool startup.
New configuration parameters:
- `workerthreads`: Defines the number of worker threads for
imtcp. If set to 1, no worker pool is created.
- `starvationProtection.maxReads`: Defines the maximum number
of consecutive reads a worker can perform before being
interrupted to allow other sessions to be processed.
New impstats counters (emitted only when `workerthreads > 1`):
- `runs`: Number of times the worker thread has been invoked.
- `read`: Number of read calls performed by the worker.
For TLS, this includes read/write calls.
- `accept`: Number of `accept()` calls handled by the worker.
- `starvation_protect`: Number of times a socket was sent
back to the queue due to reaching the maximum number of
consecutive requests, ensuring fair scheduling of sessions.
These changes significantly enhance rsyslog’s TCP handling
performance and stability, particularly in high-volume
environments.
Closes#5529, #5532, #5578, #5580.
Updated the imtcp test suite to set the "SocketBacklog" parameter to a higher
value to accommodate tests with a large number of concurrent connections.
This change addresses potential CI flakes caused by SYN queue overflows during
test runs and significantly reduces test runtime by avoiding delays during
session setup.
The adjusted SocketBacklog parameter ensures smoother handling of high
connection counts, improving test reliability and performance.
Also now permit interactivly running tests without explicitly setting
$srcdir. This now works if we are inside ./tests and fails, as before,
when we are in a different directory.
Detected by shellcheck via CodeFactor.io
Support tools (like tcpflood) are also upgraded to support the
necessary dynamic port.s
This is part of the effort to make parallel testing possible.
We move parts of the cleanup to the buildbot cleanup, as we cannot
clean out instances on each test when we run parallel tests.
changes some of the test commands to use bash functions
includes some small bug fixes to tests where bugs were
previously not seen due to different plumbing.
The imdiag module now can very effectively inject messages, which also
frees us from uncertainties of tcp reception and processing. All shell
script based tests have been modularized, what makes it far easier to
create new tests. Also, the test bench now executes more reliable and
much faster, because we can now rely on actual engine information where
we previously did just a dumb sleep.
It is checked that many tcp connections are properly handled. While
adding this test, I noticed that there is a bug in imtcp that prevents
creation of more than 200 connections. This bug still exists, so the
test suite currently fails (what is correct). Will be addressed soon.
