[patch 0/8] New RT Task Balancing

From: Steven Rostedt
Date: Fri Oct 19 2007 - 14:45:34 EST



Currently in mainline the balancing of multiple RT threads is quite broken.
That is to say that a high priority thread that is scheduled on a CPU
with a higher priority thread, may need to unnecessarily wait while it
can easily run on another CPU that's running a lower priority thread.

Balancing (or migrating) tasks in general is an art. Lots of considerations
must be taken into account. Cache lines, NUMA and more. This is true
with general processes which expect high through put and migration can
be done in batch. But when it comes to RT tasks, we really need to
put them off to a CPU that they can run on as soon as possible. Even
if it means a bit of cache line flushing.

Right now an RT task can wait several milliseconds before it gets scheduled
to run. And perhaps even longer. The migration thread is not fast enough
to take care of RT tasks.

To demonstrate this, I wrote a simple test.

http://rostedt.homelinux.com/rt/rt-migrate-test.c

(gcc -o rt-migrate-test rt-migrate-test.c -lpthread)

This test expects a parameter to pass in the number of threads to create.
If you add the '-c' option (check) it will terminate if the test fails
one of the iterations. If you add this, pass in +1 threads.

For example, on a 4 way box, I used

rt-migrate-test -c 5

What this test does is to create the number of threads specified (in this
case 5). Each thread is set as an RT FIFO task starting at a specified
prio (default 2) and each thread being one priority higher. So with this
example the 5 threads created are at priorities 2, 3, 4, 5, and 6.

The parent thread sets its priority to one higher than the highest of
the children (this example 7). It uses pthread_barrier_wait to synchronize
the threads. Then it takes a time stamp and starts all the threads.
The threads when woken up take a time stamp and compares it to the parent
thread to see how long it took to be awoken. It then runs for an
interval (20ms default) in a busy loop. The busy loop ends when it reaches
the interval delta from the start time stamp. So if it is preempted, it
may not actually run for the full interval. This is expected behavior
of the test.

The numbers recorded are the delta from the thread's time stamp from the
parent time stamp. The number of iterations it ran the busy loop for, and
the delta from a thread time stamp taken at the end of the loop to the
parent time stamp.

Sometimes a lower priority task might wake up before a higher priority,
but this is OK, as long as the higher priority process gets the CPU when
it is awoken.

At the end of the test, the iteration data is printed to stdout. If a
higher priority task had to wait for a lower one to finish running, then
this is considered a failure. Here's an example of the output from
a run against git commit 4fa4d23fa20de67df919030c1216295664866ad7.

1: 36 33 20041 39 33
len: 20036 20033 40041 20039 20033
loops: 167789 167693 227167 167829 167814

On iteration 1 (starts at 0) the third task started at 20ms after the parent
woke it up. We can see here that the first two tasks ran to completion
before the higher priority task was even able to start. That is a
20ms latency for the higher priority task!!!

So people who think that their audio would lose most latencies by upping
the priority, may be in for a surprise. Since some kernel threads (like
the migration thread itself) may cause this latency.

To solve this issue, I've changed the RT task balancing from a passive
method (migration thread) to an active method. This new method is
to actively push or pull RT tasks when they are woken up or scheduled.

On wake up of a task if it is an RT task, and there's already an RT task
of higher priority running on its runqueue, we initiate a push_rt_tasks
algorithm. This algorithm looks at the highest non-running RT task
and tries to find a CPU where it can run on. It only migrates the RT
task if it finds a CPU (of lowest priority) where the RT task
can run on and can preempt the currently running task on that CPU.
We continue pushing RT tasks until we can't push anymore.

If a RT task fails to be migrated we stop the pushing. This is possible
because we are always looking at the highest priority RT task on the
run queue. And if it can't migrate, then most likely the lower RT tasks
can not either.

There is one case that is not covered by this patch set. That is that
when the highest priority non running RT task has its CPU affinity
in such a way that it can not preempt any tasks on the CPUs running
on CPUs of its affinity. But a lower priority task has a larger affinity
to CPUs that it can run on. This is a case where the lower priority task
will not be migrated to those CPUS (although those CPUs may pull that task).
Currently this patch set ignores this scenario.

Another case where we push RT tasks is in the finish_task_switch. This is
done since the running RT task can not be migrated while it is running.
So if an RT task is preempted by a higher priority RT task, we can
migrate the RT task being preempted at that moment.

We also actively pull RT tasks. Whenever a runqueue is about to lower its
priority (schedule a lower priority task) a check is done to see if that
runqueue can pull RT tasks to it to run instead. A search is made of all
overloaded runqueues (runqueues with more than one RT task scheduled on it)
and checked to see if they have an RT task that can run on the runqueue
(affinity matches) and is of higher priority than the task the runqueue
is about to schedule. The pull algorithm pulls all RT tasks that match
this case.

With this patch set, I ran the rt-migrate-test over night in a while
loop with the -c option (which terminates upon failure) and it passed
over 6500 tests (each doing 50 wakeups each).

Here's an example of the output from the patched kernel. This is just to
explain it a bit more.

1: 20060 61 55 56 61
len: 40060 20061 20055 20056 20061
loops: 227255 146050 168104 145965 168144

2: 40 46 31 35 42
len: 20040 20046 20031 20035 20042
loops: 28 167769 167781 167668 167804

The first iteration (really 2cd, since we start at zero), is a typical run.
The lowest prio task didn't start executing until the other 4 tasks finished
and gave up the CPU.

The second iteration seems at first like a failure. But this is actually fine.
The lowest priority task just happen to schedule onto a CPU before the
higher priority tasks were woken up. But as you can see from this example,
the higher priority tasks still were able to get scheduled right away and
in doing so preempted the lower priority task. This can be seen by the
number of loops that the lower priority task was able to complete. Only 28.
This is because the busy loop terminates when the time stamp reaches the
time interval (20ms here) from the start time stamp. Since the lower priority
task was able to sneak in and start, it's time stamp was low. So after it
got preempted, and rescheduled, it was already past the run time interval
so it simply ended the loop.

Finally, the CFS RT balancing had to be removed in order for this to work.
Testing showed that the CFS RT balancing would actually pull RT tasks
from runqueues already assigned to the proper runqueues, and again cause
latencies. With this new approach, the CFS RT balancing is not needed,
and I suggest that these patches replace the current CFS RT balancing.

Also, let me stress, that I made a great attempt to have this cause
as little overhead (practically none) to the non RT cases. Most of these
algorithms only take place when more than one RT task is scheduled on the
same runqueue.

Special thanks goes to Gregory Haskins for his advice and back and forth
on IRC with ideas. Although I didn't use his actual patches (his were
against -rt) it did help me clean up some of my code. Also, thanks go to
Ingo Molnar himself for taking some ideas from his RT balance code in
the -rt patch.


Comments welcomed!

-- Steve

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/