Ø
Next question; If we decide that a sleep of less than one tick should be a
Ø
sleep of one tick, then perhaps Ralph’s issue is arguably caused by an issue
Ø
in the
posix implementation of epicsThreadSleep instead of in
epicsTimer.
Ø
Fixing only
posix timer (and any other inconsistent implementations) would
Ø
presumably be an optimal solution where Ralph’s issue is solved and also
Ø
periodic timers can remain more statistically accurate (on average)? Note
Ø
that the
epicsTimer class _is_ used to schedule the scan threads, ca beacons, etc.
BTW, it occurred to me even yesterday (and probably before) that a good idea would
be to subtract the one half of a tick from the delay, to the next timer expiration, returned
from
epicsTimer::process instead of subtracting this one half tick amount from the expiration
time stored in all of the timer objects.
This would have two beneficial impacts.
o The timer would never be allowed to expire early, as verified using
epicsTime::getCureent().
o We can still make adjustments to this returned delay amount to get the wakeup to occur as
close as possible to the correct time on average.
I was thinking about this alternative even yesterday, but there is a hitch. What does one do if
waking up between zero and one half of a tick too early that doesn’t consume
cpu or delay
timer expiration by much worse - between 1 tick <= delay <= 2 ticks.
So the bottom line is that, before sending my message today and yesterday, I considered the
above alternative and rejected it, due to concerns about wasted
cpu.
PS: Note that if we don’t allow a timer to expire as much as one half tick too early then we
will also be obliged to accept that a timers will expire on average one half of a tick too late.
Nevertheless, I am happy to implement that
alterantive if this what everyone wants.
PPS: I still think that we should make all of the
epicsThreadSleep OS specific implementations
consistent WRT sleeps requested in seconds less than one
half of a tick. Furthermore, I feel
that no one would be (should be) calling
epicsThreadSleep if they wanted no amount of slumber.
PPPS: Ralph’s issue might also be resolved by deciding _not_ to subtract out the one half tick
from the timer expiration time if the requested timer delay is less than one tick. This additional
overhead would be necessary only if the
epicsThreadSleep is allowed to interpret sleeps requested
less than one half of a tick as no sleep.
Jeff
From: Hill,
Jeff
Sent: Friday, June 08, 2012 8:24 AM
To: 'Eric Norum'; Ralph Lange
Cc: EPICS Core Talk; Dirk Zimoch
Subject: RE: epicsTimer and rounding
Ø
The posix implementation of epicsThreadSleep calls nanosleep
Ø
asking for a 0 delay, and gets a 0 delay.
This is certainly inconsistent with the windows implementation which (intends) to round
any floating point seconds sleep requested of less than one tick to a tick. Rounding sleeps
of between 0 and .5 seconds to no sleep definitely seems wrong to me. We should probably
clarify what behavior should be implemented by each OS specific stub in the application
developers guide, and then make all of them consistent?
Ø
I've never encountered a 'sleep/pause/nap/delay' function that could return early.
Next question; If we decide that a sleep of less than one tick should be a sleep of one tick, then
perhaps Ralph’s issue is arguably caused by an issue in the posix implementation of
epicsThreadSleep instead of in epicsTimer. Fixing only posix timer (and any other inconsistent
implementations) would presumably be an optimal solution where Ralph’s issue is solved and
also periodic timers can remain more statistically accurate (on average)? Note that the epicsTimer
class _is_ used to schedule the scan threads, ca beacons, etc.
I will certainly go with whatever the consensus ends up being.
Jeff
On Jun 8, 2012, at 6:35 AM, Ralph Lange wrote:
It is application dependent. Short sleeps are usually intended as waits,
so the application assumes it will wake up no earlier than after the
requested time. This is why the OS specific sleep functions always round
up, never down. And this is why I think the epicsTimer should behave the
same way.
In case an application wants a statistically optimized behavior, it can
easily get the quantum and do the subtraction itself.
I know this is symmetric (the application could add half a quantum when
it wants the round-up behavior), but I would strongly suggest that the
semantics of the epicsTimer follow the semantics of all reasonable OS's
sleep implementations, and always round up to the next quantum.
I've never encountered a 'sleep/pause/nap/delay' function that could return early.