Currently we have per-thread arrays of trees and counts, but these
ones unfortunately share cache lines and are accessed very often. This
patch moves the task-specific stuff into a structure taking a multiple
of a cache line, and has one such per thread. Just doing this has
reduced the cache miss ratio from 19.2% to 18.7% and increased the
12-thread test performance by 3%.
It starts to become visible that we really need a process-wide per-thread
storage area that would cover more than just these parts of the tasks.
The code was arranged so that it's easy to move the pieces elsewhere if
needed.
Now we still have a main contention point with the timers in the main
wait queue, but the vast majority of the tasks are pinned to a single
thread. This patch creates a per-thread wait queue and queues a task
to the local wait queue without any locking if the task is bound to a
single thread (the current one) otherwise to the shared queue using
locking. This significantly reduces contention on the wait queue. A
test with 12 threads showed 11 ms spent in the WQ lock compared to
4.7 seconds in the same test without this change. The cache miss ratio
decreased from 19.7% to 19.2% on the 12-thread test, and its performance
increased by 1.5%.
Another indirect benefit is that the average queue size is divided
by the number of threads, which roughly removes log(nbthreads) levels
in the tree and further speeds up lookups.
In tasklet_free(), if we're currently in the runnable task list, don't
forget to decrement taks_list_size, or it'll end up being to big, and we may
not process tasks in the global runqueue.
Set the flag for the current thread in active_threads_mask when waking a
tasklet, or we will never run it if no tasks are available.
This is 1.9-specific, no backport is needed.
Modify tasklet_wakeup() so that it handles a task as well, and inserts it
directly into the tasklet list, making it effectively a tasklet.
This should make future developments easier.
As __task_wakeup() is responsible for increasing
rqueue_local[tid]/global_rqueue_size, make __task_unlink_rq responsible for
decreasing it, as process_runnable_tasks() isn't the only one that removes
tasks from runqueues.
Up until now, a tasklet couldn't be free'd while it was in the list, it is
no longer the case, so make sure we remove it from the list before freeing it.
To do so, we have to make sure we correctly initialize it, so use LIST_INIT,
instead of setting the pointers to NULL.
To make sure we don't inadvertently insert task in the global runqueue,
while only the local runqueue is used without threads, make its definition
and usage conditional on USE_THREAD.
When building without threads enabled, instead of just using the global
runqueue, just use the local runqueue associated with the only thread, as
that's what is now expected for a single thread in prcoess_runnable_tasks().
This should fix haproxy when built without threads.
This function returns true is some notifications are registered.
This function is usefull for the following patch
BUG/MEDIUM: lua/socket: Sheduling error on write: may dead-lock
It should be backported in 1.6, 1.7 and 1.8
Don't forget to increase tasks_run_queue when we're adding a task to the
tasklet list, and to decrease it when we remove a task from a runqueue,
or its value won't be accurate, and could lead to tasks not being executed
when put in the global run queue.
1.9-dev only, no backport is needed.
Introduce tasklets, lightweight tasks. They have no notion of priority,
they are just run as soon as possible, and will probably be used for I/O
later.
For the moment they're used to replace the temporary thread-local list
that was used in the scheduler. The first part of the struct is common
with tasks so that tasks can be cast to tasklets and queued in this list.
Once a task is in the tasklet list, it has its leaf_p set to 0x1 so that
it cannot accidently be confused as not in the queue.
Pure tasklets are identifiable by their nice value of -32768 (which is
normally not possible).
A lot of tasks are run on one thread only, so instead of having them all
in the global runqueue, create a per-thread runqueue which doesn't require
any locking, and add all tasks belonging to only one thread to the
corresponding runqueue.
The global runqueue is still used for non-local tasks, and is visited
by each thread when checking its own runqueue. The nice parameter is
thus used both in the global runqueue and in the local ones. The rare
tasks that are bound to multiple threads will have their nice value
used twice (once for the global queue, once for the thread-local one).
While running a task, we may try to delete and free a task that is about to
be run, because it's part of the local tasks list, or because rq_next points
to it.
So flag any task that is in the local tasks list to be deleted, instead of
run, by setting t->process to NULL, and re-make rq_next a global,
thread-local variable, that is modified if we attempt to delete that task.
Many thanks to PiBa-NL for reporting this and analysing the problem.
This should be backported to 1.8.
The thread patches adds refcount for notifications. The notifications are
used with the Lua cosocket. These refcount free the notifications when
the session is cleared. In the Lua task case, it not have sessions, so
the nofications are never cleraed.
This patch adds a garbage collector for signals. The garbage collector
just clean the notifications for which the end point is disconnected.
This patch should be backported in 1.8
During the migration to the second version of the pools, the new
functions and pool pointers were all called "pool_something2()" and
"pool2_something". Now there's no more pool v1 code and it's a real
pain to still have to deal with this. Let's clean this up now by
removing the "2" everywhere, and by renaming the pool heads
"pool_head_something".
a bitfield has been added to know if there are runnable tasks for a thread. When
a task is woken up, the bits corresponding to its thread_mask are set. When all
tasks for a thread have been evaluated without any wakeup, the thread is removed
from active ones by unsetting its tid_bit from the bitfield.
This macro should be used to declare variables or struct members depending on
the USE_THREAD compile option. It avoids the encapsulation of such declarations
between #ifdef/#endif. It is used to declare all lock variables.
Currently the task scheduler suffers from an O(n) lookup when
skipping tasks that are not for the current thread. The reason
is that eb32_lookup_ge() has no information about the current
thread so it always revisits many tasks for other threads before
finding its own tasks.
This is particularly visible with HTTP/2 since the number of
concurrent streams created at once causes long series of tasks
for the same stream in the scheduler. With only 10 connections
and 100 streams each, by running on two threads, the performance
drops from 640kreq/s to 11.2kreq/s! Lookup metrics show that for
only 200000 task lookups, 430 million skips had to be performed,
which means that on average, each lookup leads to 2150 nodes to
be visited.
This commit backports the principle of scope lookups for ebtrees
from the ebtree_v7 development tree. The idea is that each node
contains a mask indicating the union of the scopes for the nodes
below it, which is fed during insertion, and used during lookups.
Then during lookups, branches that do not contain any leaf matching
the requested scope are simply ignored. This perfectly matches a
thread mask, allowing a thread to only extract the tasks it cares
about from the run queue, and to always find them in O(log(n))
instead of O(n). Thus the scheduler uses tid_bit and
task->thread_mask as the ebtree scope here.
Doing this has recovered most of the performance, as can be seen on
the test below with two threads, 10 connections, 100 streams each,
and 1 million requests total :
Before After Gain
test duration : 89.6s 4.73s x19
HTTP requests/s (DEBUG) : 11200 211300 x19
HTTP requests/s (PROD) : 15900 447000 x28
spin_lock time : 85.2s 0.46s /185
time per lookup : 13us 40ns /325
Even when going to 6 threads (on 3 hyperthreaded CPU cores), the
performance stays around 284000 req/s, showing that the contention
is much lower.
A test showed that there's no benefit in using this for the wait queue
though.
It was a leftover from the last cleaning session; this mask applies
to threads and calling it process_mask is a bit confusing. It's the
same in fd, task and applets.
2 global locks have been added to protect, respectively, the run queue and the
wait queue. And a process mask has been added on each task. Like for FDs, this
mask is used to know which threads are allowed to process a task.
For many tasks, all threads are granted. And this must be your first intension
when you create a new task, else you have a good reason to make a task sticky on
some threads. This is then the responsibility to the process callback to lock
what have to be locked in the task context.
Nevertheless, all tasks linked to a session must be sticky on the thread
creating the session. It is important that I/O handlers processing session FDs
and these tasks run on the same thread to avoid conflicts.
These notification management function and structs are generic and
it will be better to move in common parts.
The notification management functions and structs have names
containing some "lua" references because it was written for
the Lua. This patch removes also these references.
task_init() is called exclusively by task_new() which is the only way
to create a task. Most callers set t->expire to TICK_ETERNITY, some set
it to another value and a few like Lua don't set it at all as they don't
need a timeout, causing random values to be used in case the task gets
queued.
Let's always set t->expire to TICK_ETERNITY in task_init() so that all
tasks are now initialized in a clean state.
This patch can be backported as it will definitely make the code more
robust (at least the Lua code, possibly other places).
In order to authorize call of task_wakeup on running task:
- from within the task handler itself.
- in futur, from another thread.
The lookups on runqueue and waitqueue are re-worked
to prepare multithread stuff.
If task_wakeup is called on a running task, the woken
message flags are savec in the 'pending_state' attribute of
the state. The real wakeup is postponed at the end of the handler
process and the woken messages are copied from pending_state
to the state attribute of the task.
It's important to note that this change will cause a very minor
(though measurable) performance loss but it is necessary to make
forward progress on a multi-threaded scheduler. Most users won't
ever notice.
<run_queue> is used to track the number of task in the run queue and
<run_queue_cur> is a copy used for the reporting purpose. These counters has
been renamed, respectively, <tasks_run_queue> and <tasks_run_queue_cur>. So the
naming is consistent between tasks and applets.
[wt: needed for next fixes, backport to 1.7 and 1.6]
Actually, HAProxy uses the function "process_runnable_tasks" and
"wake_expired_tasks" to get the next task which can expires.
If a task is added with "task_schedule" or other method during
the execution of an other task, the expiration of this new task
is not taken into account, and the execution of this task can be
too late.
Actualy, HAProxy seems to be no sensitive to this bug.
This fix moves the call to process_runnable_tasks() before the timeout
calculation and ensures that all wakeups are processed together. Only
wake_expired_tasks() needs to return a timeout now.
As found by Thierry Fournier, if a task manages to kill another one and
if this other task is the next one in the run queue, we can do whatever
including crashing, because the scheduler restarts from the saved next
task. For now, there is no such concept of a task killing another one,
but with Lua it will come.
A solution consists in always performing the lookup of the first task in
the scheduler's loop, but it's expensive and costs around 2% of the
performance.
Another solution consists in keeping a global next run queue node and
ensuring that when this task gets removed, it updates this pointer to
the next one. This allows to simplify the code a bit and in the end to
slightly increase the performance (0.3-0.5%). The mechanism might still
be usable if we later migrate to a multi-threaded scheduler.
When we're stopping, we're not going to create new tasks anymore, so
let's release the task pool upon each task_free() in order to reduce
memory fragmentation.
This function is finally not needed anymore, as it has been replaced with
a per-proxy task that is scheduled when some limits are encountered on
incoming connections or when the process is stopping. The savings should
be noticeable on configs with a large number of proxies. The most important
point is that the rate limiting is now enforced in a clean and solid way.
The two new functions below make it possible to register any number
of functions or tasks to a system signal. They will be called in the
registration order when the signal is received.
struct sig_handler *signal_register_fct(int sig, void (*fct)(struct sig_handler *), int arg);
struct sig_handler *signal_register_task(int sig, struct task *task, int reason);
All files referencing the previous ebtree code were changed to point
to the new one in the ebtree directory. A makefile variable (EBTREE_DIR)
is also available to use files from another directory.
The ability to build the libebtree library temporarily remains disabled
because it can have an impact on some existing toolchains and does not
appear worth it in the medium term if we add support for multi-criteria
stickiness for instance.
It's sometimes useful at least for statistics to keep a task count.
It's easy to do by forcing the rare task creators to always use the
same functions to create/destroy a task.
The top of a duplicate tree is not where bit == -1 but at the most
negative bit. This was causing tasks to be queued in reverse order
within duplicates. While this is not dramatic, it's incorrect and
might lead to longer than expected duplicate depths under some
circumstances.
Since we're now able to search from a precise expiration date in
the timer tree using ebtree 4.1, we don't need to maintain 4 trees
anymore. Not only does this simplify the code a lot, but it also
ensures that we can always look 24 days back and ahead, which
doubles the ability of the previous scheduler. Indeed, while based
on absolute values, the timer tree is now relative to <now> as we
can always search from <now>-31 bits.
The run queue uses the exact same principle now, and is now simpler
and a bit faster to process. With these changes alone, an overall
0.5% performance gain was observed.
Tests were performed on the few wrapping cases and everything works
as expected.
Most of the time, task_queue() will immediately return. By extracting
the preliminary checks and putting them in an inline function, we can
significantly reduce the number of calls to the function itself, and
most of the tests can be optimized away due to the caller's context.
Another minor improvement in process_runnable_tasks() consisted in
taking benefit from the processor's branch prediction unit by making
a special case of the process_session() callback which is by far the
most common one.
All this improved performance by about 1%, mainly during the call
from process_runnable_tasks().
Timers are unsigned and used as tree positions. Ticks are signed and
used as absolute date within current time frame. While the two are
normally equal (except zero), it's important not to confuse them in
the code as they are not interchangeable.
We add two inline functions to turn each one into the other.
The comments have also been moved to the proper location, as it was
not easy to understand what was a tick and what was a timer unit.
In many situations, we wake a task on an I/O event, then queue it
exactly where it was. This is a real waste because we delete/insert
tasks into the wait queue for nothing. The only reason for this is
that there was only one tree node in the task struct.
By adding another tree node, we can have one tree for the timers
(wait queue) and one tree for the priority (run queue). That way,
we can have a task both in the run queue and wait queue at the
same time. The wait queue now really holds timers, which is what
it was designed for.
The net gain is at least 1 delete/insert cycle per session, and up
to 2-3 depending on the workload, since we save one cycle each time
the expiration date is not changed during a wake up.
It's very frequent to require some information about the
reason why a task is running. Some flags have been added
so that a task now knows if it got woken up due to I/O
completion, timeout, etc...
A test has shown that more than 16% of the calls to task_wakeup()
could be avoided because the task is already woken up. So make it
inline and move the test to the inline part.
