[erlang-questions] linked-in driver blocks net_kernel tick?

[Wallentin Dahlberg]

I think his argument was that a driver or nif that does not use an
async-thread for potentially blocking calls is a seriously broken driver.
Consider the non-smp case. It will halt the beam and hinder important
processes to be scheduled. I agree that in the smp case, one scheduler
should not block the other schedulers in damaging calls. If a developer
wants to destroy a scheduler with a broken driver, he should be free to do
so.

This was the fear with nifs. With nifs developers has an easy tool to really
destroy the system in order to "increase performance" and implement 3rd
party libs. There are several cases with different impact,
  1) destroy soft-real-time properties - reduction count badness
  2) destroy concurrency with blocking calls - scheduler badness
  3) destroy the system with faulty drivers (seg fault) - pure badness

Some of these issues can be mitigated if the developer implements async
threads, i.e. schedules operations to the async-pool.

I feel that this is not ideal and is a heritage of ancient times.

The problem in this case is that time does not progress in the system. Time
is measured in reductions and each call is a reduction. At least this is the
case with normal code. There are som special cases too, for instance a
message sent "bumbs" the reduction count of the sender. Since native code
(nif, bifs and drivers) do not increase reductions during its call but
instead penalize the process after the call, time does not progress during
the execution (as opposed to beam code). When a process reaches the
reduction-limit it is scheduled out. Why reduction counters instead of time
slices? Supposedly it much faster (according to Björn). It is a design
decision with trade-offs. The solution is fast and nimble, it has certain
characteristics that are favorable and has some characteristics that are
less favorable. I would favor time-slices since i think it would be fairer
to the system and potentially we could save a register. Exactly how it
should be done is a question for a different time.

The load balancing in the scheduler is checked when a certain reduction
count is reached for that scheduler. We do not want to check this too often
since it will then become a serialization point.

But fear not, there is a (beautiful) solution that is being discussed in the
erts-team. Hopefully we can agree on the details.

// Björn-Egil


2011/1/22 Dan Gudmundsson <dgud@REDACTED>

> Rickard who have implemented this should explain it.
>
> If I have understood it correctly, it works like this:
> If a scheduler do not have any work to do it will be disabled.
> It will be disabled until a live thread discovers it have to much work and
> wakes a sleeping scheduler. The run-queues are only checked when processes
> are scheduled.
>
> Since in this case the only living scheduler is busy for a very long time,
> no queue checking will be done and the all schedulers will be blocked until
> the call to the driver is complete.
>
> We had a long discussion during lunch about it, and we didn't agree
> how it should
> work. :-)
>
> I agree that zlib is broken and it should be fixed but I still believe that
> it
> breaks the rule about least astonishment, if I have 16 schedulers and
> one is blocked
> in a long function call I still expect other code to be invoked.
> Rickards thought is that
> such call should never happen and should be called through an async
> driver or a separate
> thread. I guess it will take a couple of more lunches to come to a
> conclusion :-)
>
> /Dan
>
> On Fri, Jan 21, 2011 at 10:25 PM, Ryan Zezeski <rzezeski@REDACTED> wrote:
> > Dan,
> >
> > Thanks for the reply, I'll be sure to chunk my data.  I was using the
> gzip/1
> > call for convenience.
> >
> > That said, I'm still a little fuzzy on something you said.  Why is it
> that
> > the "distribution" process is scheduled on the same scheduler that's
> running
> > the call to the driver?  Why not schedule it on one of the 15 other
> > schedulers that are currently sleeping?  Does this mean any other message
> I
> > send will also be blocked?  Dare I ask, how does the scheduling work
> > exactly?
> >
> > -Ryan
> >
> > On Fri, Jan 21, 2011 at 5:16 AM, Dan Gudmundsson <dgud@REDACTED>
> wrote:
> >
> >> All c-calls blocks a schedulers, if they are not pushed out to a thread.
> >>
> >> In this case it's a bug in the zlib module (probably by me) gzip should
> >> chunk up the input before invoking the driver.
> >>
> >> What happens is that all schedulers go to sleep because there is no work
> to
> >> do,
> >> except the one invoking the driver, a ping is received and wakes up
> >> the "distribution" process
> >> which gets queued up on only scheduler that is awake, but that
> >> scheduler is blocked
> >> in an "eternal" call. The pings never become processed and the
> >> distributions times out.
> >>
> >> You can wait for a patch or use zlib api to chunk up compression your
> self,
> >> see
> >> implementation of gzip in zlib module.
> >>
> >> /Dan
> >>
> >> On Fri, Jan 21, 2011 at 2:48 AM, Ryan Zezeski <rzezeski@REDACTED>
> wrote:
> >> > So...can anyone explain to me why zlib:gzip/1 is causing the
> net_kernel
> >> tick
> >> > to be blocked?  Do linked-in drivers block it's scheduler like NIFs?
>  I'm
> >> > really curious on this one :)
> >> >
> >> > -Ryan
> >> >
> >> > On Tue, Jan 18, 2011 at 6:53 PM, Ryan Zezeski <rzezeski@REDACTED>
> >> wrote:
> >> >
> >> >> Apologies, the example I copied was run on my mac.
> >> >>
> >> >> This is what I have on the actual production machine:
> >> >>
> >> >> Erlang R14A (erts-5.8) [source] [64-bit] [smp:16:16] [rq:16]
> >> >> [async-threads:0] [hipe] [kernel-poll:false]
> >> >>
> >> >> To be certain, I ran the same example (except this time using two
> >> physical
> >> >> machines) and achieved the same result.  Namely, the 'bar' node
> claims
> >> 'foo'
> >> >> is not responding and thus closes the connection.  Whatever this is,
> >> I've
> >> >> now easily reproduced it on two different OSs, with 2 different
> Erlang
> >> >> versions.
> >> >>
> >> >> -Ryan
> >> >>
> >> >> On Tue, Jan 18, 2011 at 6:04 PM, Alain O'Dea <alain.odea@REDACTED>
> >> wrote:
> >> >>
> >> >>> On 2011-01-18, at 18:54, Ryan Zezeski <rzezeski@REDACTED> wrote:
> >> >>>
> >> >>> > Hi everyone,
> >> >>> >
> >> >>> > Some of you may remember my latest question where I was having
> weird
> >> >>> node
> >> >>> > timeout issues that I couldn't explain and I thought it might be
> >> related
> >> >>> to
> >> >>> > the messages I was passing between my nodes.  Well, I pinpointed
> the
> >> >>> problem
> >> >>> > to a call to zlib:gzip/1.  At first I was really surprised by
> this,
> >> as
> >> >>> such
> >> >>> > a harmless line of code surely should have nothing to do with the
> >> >>> ability
> >> >>> > for my nodes to communicate.  However, as I dug further I realized
> >> gzip
> >> >>> was
> >> >>> > implemented as a linked-in driver and I remember reading things
> about
> >> >>> how
> >> >>> > one has to take care with them because they can trash the VM with
> >> them.
> >> >>>  I
> >> >>> > don't remember reading anything about them blocking code, and even
> if
> >> >>> they
> >> >>> > do I fail to see why my SMP enabled node (16 cores) would allow
> this
> >> one
> >> >>> > thread to block the tick.  It occurred to me that maybe the
> scheduler
> >> >>> > responsible for that process is the one blocked by the driver.  Do
> >> >>> processes
> >> >>> > have scheduler affinity?  That would make sense, I guess.
> >> >>> >
> >> >>> > I've "fixed" this problem simply by using a plain port (i.e. run
> in
> >> it's
> >> >>> own
> >> >>> > OS process).  For my purposes, this actually makes more sense in
> the
> >> >>> > majority of the places I was making use of gzip.  Can someone
> >> enlighten
> >> >>> me
> >> >>> > as to exactly what is happening behind the scenes?
> >> >>> >
> >> >>> > To reproduce I create a random 1.3GB file:
> >> >>> >
> >> >>> > dd if=/dev/urandom of=rand bs=1048576 count=1365
> >> >>> >
> >> >>> > Then start two named nodes 'foo' and 'bar', connect them, read in
> the
> >> >>> file,
> >> >>> > and then compress said file.  Sometime later (I think around 60+
> >> >>> seconds)
> >> >>> > the node 'bar' will claim that 'foo' is not responding.
> >> >>> >
> >> >>> > [progski@REDACTED ~/tmp_code/node_timeout] erl -name foo
> >> >>> > Erlang R14B (erts-5.8.1) [source] [64-bit] [smp:2:2] [rq:2]
> >> >>>
> >> >>> Your SMP node seems to be capped at smp:2:2 when it out to be
> smp:16.
> >> >>>  Some resource limit may be holding back the system. That said zlib
> >> should
> >> >>> not ever cause this issue.
> >> >>>
> >> >>> > [async-threads:0] [hipe] [kernel-poll:false]
> >> >>> >
> >> >>> > Eshell V5.8.1  (abort with ^G)
> >> >>> > (foo@REDACTED)1> net_adm:ping('bar@REDACTED').
> >> >>> > pong
> >> >>> > (foo@REDACTED)2> nodes().
> >> >>> > ['bar@REDACTED']
> >> >>> > (foo@REDACTED)3> {ok,Data} = file:read_file("rand").
> >> >>> > {ok,<<103,5,115,210,177,147,53,45,250,182,51,32,250,233,
> >> >>> >      39,253,102,61,73,242,18,159,45,185,232,80,33,...>>}
> >> >>> > (foo@REDACTED)4> zlib:gzip(Data).
> >> >>> > <<31,139,8,0,0,0,0,0,0,3,0,15,64,240,191,103,5,115,210,
> >> >>> >  177,147,53,45,250,182,51,32,250,233,...>>
> >> >>> > (foo@REDACTED)5>
> >> >>> >
> >> >>> >
> >> >>> > [progski@REDACTED ~/tmp_code/node_timeout] erl -name bar
> >> >>> > Erlang R14B (erts-5.8.1) [source] [64-bit] [smp:2:2] [rq:2]
> >> >>> > [async-threads:0] [hipe] [kernel-poll:false]
> >> >>> >
> >> >>> > Eshell V5.8.1  (abort with ^G)
> >> >>> > (bar@REDACTED)1> nodes().
> >> >>> > ['foo@REDACTED']
> >> >>> > (bar@REDACTED)2>
> >> >>> > =ERROR REPORT==== 18-Jan-2011::17:16:10 ===
> >> >>> > ** Node 'foo@REDACTED' not responding **
> >> >>> > ** Removing (timedout) connection **
> >> >>> >
> >> >>> >
> >> >>> > Thanks,
> >> >>> >
> >> >>> > -Ryan
> >> >>>
> >> >>
> >> >>
> >> >
> >>
> >
>
> ________________________________________________________________
> erlang-questions (at) erlang.org mailing list.
> See http://www.erlang.org/faq.html
> To unsubscribe; mailto:erlang-questions-unsubscribe@REDACTED
>
>