more data: SCHED_ULE+PREEMPTION is the problem

Stefan Esser se at
Mon Apr 9 10:07:52 UTC 2018

Am 07.04.18 um 16:18 schrieb Peter:
> 3. kern.sched.preempt_thresh
> I could make the problem disappear by changing kern.sched.preempt_thresh  from
> the default 80 to either 11 (i5-3570T) or 7 (p3) or smaller. This seems to
> correspond to the disk interrupt threads, which run at intr:12 (i5-3570T) or
> intr:8 (p3).

[CC added to include Jeff as the author of the ULE scheduler ...]

Since I had somewhat similar problems on my systems (with 4 Quad-Core with SMT
enabled, i.e. 8 threads of execution) with compute bound processes keeping I/O
intensive processes from running (load average of 12 with 8 "CPUs"), and these
problems where affected by preempt_thresh, I checked how this variable is used
in the scheduler. The code is in /sys/kern/sched_ule.c.

It controls, whether a thread that has become runnable (e.g., after waiting
for disk I/O to complete) will preempt the thread currently running on "this"
CPU (i.e. the one executing this test in the kernel).

IMHO, sched_preempt should default to a much higher number than 80 (e.g. 190),
but maybe I misunderstand some of the details ...

static inline int
sched_shouldpreempt(int pri, int cpri, int remote)

The parameters are:

	pri: the priority if the now runnable thread
	cpri: the priority of the thread that currently runs on "this" CPU
	remote: whether to consider preempting a thread on another CPU

The priority values are those displayed by top or ps -l as "PRI", but with an
offset of 100 applied (i.e. pri=120 is displayed as PRI=20 in top).

If this thread has less priority than the currently executing one (cpri), the
currently running thread will not be preempted:


     * If the new priority is not better than the current priority there is

     * nothing to do.

    if (pri >= cpri)
        return (0);

If the current thread is the idle thread, it will always be preempted by the
now runnable thread:


     * Always preempt idle.

    if (cpri >= PRI_MIN_IDLE)
        return (1);

A value of preempt_thresh=0 (e.g. if "options PREEMPTION" is missing in the
kernel config) lets the previously running thread continue (except if was the
idle thread, which has been dealt with above). The compute bound thread may
continue until its quantum has expired.


     * If preemption is disabled don't preempt others.

    if (preempt_thresh == 0)
        return (0);

For any other value of preempt_thresh, the new priority of the thread that
just has become runnable will be compared to preempt_thresh and if this new
priority is higher (lower numeric value) or equal to preempt_thresh, the
thread for which (e.g.) disk I/O finished will preempt the current thread:


     * Preempt if we exceed the threshold.

    if (pri <= preempt_thresh)
        return (1);

===> This is the only condition that depends on preempt_thresh > 0 <===

The flag "remote" controls whether this thread will be scheduled to run, if
its priority is higher or equal to PRI_MAX_INTERACT (less than or equal to
151) and if the opposite is true for the currently running thread (cpri).
The value of remote will always be 0 on kernels built without "options SMP".


     * If we're interactive or better and there is non-interactive

     * or worse running preempt only remote processors.

    if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT)
        return (1);

The critical use of preempt_thresh is marked above. If it is 0, no preemption
will occur. On a single processor system, this should allow the CPU bound
thread to run for as long its quantum lasts.

A value of 120 (corresponding to PRI=20 in top) will allow the I/O bound
thread to preempt any other thread with lower priority (cpri > pri). But in
case of a high priority kernel thread being active during this test (with a
low numeric cpri value), the I/O bound process will not preempt that higher
priority thread (i.e. some high priority kernel thread).

Whether the I/O bound thread will run (instead of the compute bound) after
the higher priority thread has given up the CPU, will depend on the scheduler
decision which thread to select. And for "timeshare" threads, this will often
not be the higher priority (I/O bound) thread, but the compute bound thread,
which then may execute until next being interrupted by the I/O bound thread
(which will not happen, if no new I/O has been requested).

This might explain, why setting preempt_thresh to a very low value (in the
range of real-time kernel threads) enforces preemption of the CPU bound
thread, while any higher (numeric) value of preempt_thresh prevents this
and makes tdq_choose() often select the low priority CPU bound over the
higher priority I/O bound thread.

BUT the first test in sched_shouldpreempt() should prevent any user process
from ever preempting a real-time thread "if (pri >= cpri) return 0;".

For preemption to occur,  pri must be numerically lower than cpri, and
pri must be numerically lower than or equal to preempt_thresh.

> a. with kern.sched.preempt_thresh=80
> $ lz4 DATABASE_TEST_FILE /dev/null & while true;
>   do ps -o pid,pri,"%cpu",command -p 2119,$!
>   sleep 3
> done
> [1] 6073
> 6073  52  6.5 lz4 DATABASE_TEST_FILE /dev/null
> 2119  99 91.5 -bash (bash)

The I/O bound thread does not preempt the compute bound thread, when becoming
runnable (data arrived from disk).

With the value of preempt_thresh=80 (corresponding to PRI=-20) only real-time
threads may cause preemption, the I/O bound thread can not (PRI=52 / pri=152).

A value of preempt_thresh in the range of 190 (corresponding to PRI=90) should
allow the lz4 process to preempt the CPU bound process (with higher pri/PRI).

> b. with kern.sched.preempt_thresh=11
> 4920  21  0.0 lz4 DATABASE_TEST_FILE /dev/null
> 2119 101 93.5 -bash (bash)
> 4920  85 43.0 lz4 DATABASE_TEST_FILE /dev/null
> 2119  85 45.5 -bash (bash)

Such a low preempt_thresh does not allow any user process to preempt any other
one (except when running with temporarily increased priority in the kernel).

Only a kernel thread (soft interrupt?) at might cause preemption, and if the
interrupt is due to a read issued by the I/O bound thread completing, then the
I/O bound process is not the one being preempted. This will make the timeshare
scheduler select the process with higher priority (lower PRI) that did not
recently run (i.e. the I/O bound process, if both have the same PRI), when the
kernel thread goes to sleep.

But (if my analysis is correct) this indicates, that preempt_thresh set to an
extremely low value just helps by accident. The kernel thread interrupts the
CPU bound thread, and the I/O bound thread is selected as the next runnable
thread in the time-share run queue, either because of its lower PRI value or
because it did not run last before the preemption occurred (with equal PRI
for both).

But, in fact, the same scheduler selection should have occured in test (a),
too, if e.g. a soft interrupt preempts the compute bound thread. Not sure,
why this does not happen ... (And this may be an indication, that I do not
fully understand what's going on ;-) ...)

> From this we can see that in case b. both processes balance out nicely and
> meet at equal CPU shares.
> Whereas in case a., after about 10 Seconds (the first 3 records) they move to
> opposite ends of the scale and stay there.
> From this I might suppose that here is some kind of mis-calculation or
> mis-adjustment of the task priorities happening.

I'd be interested in your results with preempt_thresh set to a value of e.g.
190. The PRI=85 values in your test case (b) correspond to pri=185, and with
preempt_thresh slightly higher that that, the lz4 process should still get a
50% share of the CPU. (If its PRI grows over that of the CPU bound process,
it will not be able to preempt it, so its PRI should match the one of the CPU
bound process).

Regards, STefan

More information about the freebsd-stable mailing list