Thread behavior in 7.8.3

[John Lato]

I guess I should explain what that flag does...

The GHC RTS maintains capabilities, the number of capabilities is specified
by the `+RTS -N` option.  Each capability is a virtual machine that
executes Haskell code, and maintains its own runqueue of threads to process.

A capability will perform a context switch at the next heap block
allocation (every 4k of allocation) after the timer expires.  The timer
defaults to 20ms, and can be set by the -C flag.  Capabilities perform
context switches in other circumstances as well, such as when a thread
yields or blocks.

My guess is that either the context switching logic changed in ghc-7.8, or
possibly your code used to trigger a switch via some other mechanism (stack
overflow or something maybe?), but is optimized differently now so instead
it needs to wait for the timer to expire.

The problem we had was that a time-sensitive thread was getting scheduled
on the same capability as a long-running non-yielding thread, so the
time-sensitive thread had to wait for a context switch timeout (even though
there were free cores available!).  I expect even with -N4 you'll still see
occasional delays (perhaps <5% of calls).

We've solved our problem with judicious use of `forkOn`, but that won't
help at N1.

We did see this behavior in 7.6, but it's definitely worse in 7.8.

Incidentally, has there been any interest in a work-stealing scheduler?
There was a discussion from about 2 years ago, in which Simon Marlow noted
it might be tricky, but it would definitely help in situations like this.

John L.

On Thu, Oct 30, 2014 at 8:02 AM, Michael Jones <mike at proclivis.com> wrote:

> John,
>
> Adding -C0.005 makes it much better. Using -C0.001 makes it behave more
> like -N4.
>
> Thanks. This saves my project, as I need to deploy on a single core Atom
> and was stuck.
>
> Mike
>
> On Oct 29, 2014, at 5:12 PM, John Lato <jwlato at gmail.com> wrote:
>
> By any chance do the delays get shorter if you run your program with `+RTS
> -C0.005` ?  If so, I suspect you're having a problem very similar to one
> that we had with ghc-7.8 (7.6 too, but it's worse on ghc-7.8 for some
> reason), involving possible misbehavior of the thread scheduler.
>
> On Wed, Oct 29, 2014 at 2:18 PM, Michael Jones <mike at proclivis.com> wrote:
>
>> I have a general question about thread behavior in 7.8.3 vs 7.6.X
>>
>> I moved from 7.6 to 7.8 and my application behaves very differently. I
>> have three threads, an application thread that plots data with wxhaskell or
>> sends it over a network (depends on settings), a thread doing usb bulk
>> writes, and a thread doing usb bulk reads. Data is moved around with TChan,
>> and TVar is used for coordination.
>>
>> When the application was compiled with 7.6, my stream of usb traffic was
>> smooth. With 7.8, there are lots of delays where nothing seems to be
>> running. These delays are up to 40ms, whereas with 7.6 delays were a 1ms or
>> so.
>>
>> When I add -N2 or -N4, the 7.8 program runs fine. But on 7.6 it runs fine
>> without with -N2/4.
>>
>> The program is compiled -O2 with profiling. The -N2/4 version uses more
>> memory,  but in both cases with 7.8 and with 7.6 there is no space leak.
>>
>> I tired to compile and use -ls so I could take a look with threadscope,
>> but the application hangs and writes no data to the file. The CPU fans run
>> wild like it is in an infinite loop. It at least pops an unpainted
>> wxhaskell window, so it got partially running.
>>
>> One of my libraries uses option -fsimpl-tick-factor=200 to get around the
>> compiler.
>>
>> What do I need to know about changes to threading and event logging
>> between 7.6 and 7.8? Is there some general documentation somewhere that
>> might help?
>>
>> I am on Ubuntu 14.04 LTS. I downloaded the 7.8 tool chain tar ball and
>> installed myself, after removing 7.6 with apt-get.
>>
>> Any hints appreciated.
>>
>> Mike
>>
>>
>> _______________________________________________
>> Glasgow-haskell-users mailing list
>> Glasgow-haskell-users at haskell.org
>> http://www.haskell.org/mailman/listinfo/glasgow-haskell-users
>>
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.haskell.org/pipermail/glasgow-haskell-users/attachments/20141030/c484a7ba/attachment.html>