[argobots-discuss] modifying scheduler event frequency?

Iwasaki, Shintaro siwasaki at anl.gov
Wed Apr 14 15:55:02 CDT 2021

Hi Phil,

Thanks. I can understand a bigger picture.

> ABT_info_print_thread_stacks_in_pool()
I hope it works. Note that print_thread_stacks_in_pool() is not async-signal safe (ABT_info_trigger_print_all_thread_stacks() is an exception), so please don't call it in a signal handler.

> We use argobots almost exclusively with spack at this point.
Many HPC users use Spack to build dependent libraries. I will add some debug options (including libunwind, stack guard, ...) as well as other major options to the Spack Argobots package. We are also implementing an mprotect-based stack guard option (which is not in Argobots 1.1, though).

Overall, please give us a week or so in total.

There is large room for improvement of the debugging/profiling capability.
If you have any questions, requests, and/or suggestions, please feel free to tell us.


From: Carns, Philip H. <carns at mcs.anl.gov>
Sent: Wednesday, April 14, 2021 3:42 PM
To: Iwasaki, Shintaro <siwasaki at anl.gov>; discuss at lists.argobots.org <discuss at lists.argobots.org>
Subject: Re: [argobots-discuss] modifying scheduler event frequency?

Ah, thanks for the thorough information as always Shintaro :)

print_all_thread_stacks() was  tempting because it would potentially encompass more (in the Mochi use case, it would pick up hypothetical pools created by higher level components that we don't have a reference to).  Based on the information in this email thread, though, I think I'm better off focusing on pools under our control so that I can use print_thread_stacks_in_pool().  This should work fine; I was just over-thinking the use case.  The pools are under our own control in the vast majority of configurations.

In the big picture, I was exploring this because of a bug report we have from one of our collaborators who is getting a nonsensical hang in a complex scenario that we can't easily reproduce or attach a debugger to.  I would like to be able to send an RPC to a process at an arbitrary point in time and dump what it is up to so that we can understand why it didn't complete something it was trying to do.

libunwind sounds great :)  I probably would have been asking about that next.

I guess I'll use this as an opportunity to request/suggest that the libunwind capability be added as a variant to the argobots spack package (along with a way to enable future mprotect / stack canary checks).

We use argobots almost exclusively with spack at this point.  Not that argobots itself is hard to compile manually, but it is often one of a large number of dependencies that we need to build, so it's best to just unify them in one packaging system.  It would be straightforward for us to set up an alternative environment yaml with various argobots debugging capabilities enabled for development/debugging purposes.



On 4/14/21 3:57 PM, Iwasaki, Shintaro wrote:
Hi Phil,

Thanks for using Argobots!  The following is my answers to your questions in addition to some tips.
We would appreciate it if you could share more information about your workload and the purpose so that we can give you more specific suggestions. Also, we welcome any feature requests and bug reports.

1. How to change a scheduler's event frequency?
1.1. Predefined scheduler
First, there is no way to dynamically change the event frequency (even if you hack ABT_sched or a pointer you used in ABT_sched_get_data()... since event_freq is loaded to a local variable).
Currently, using a special ABT_sched_config when you create a scheduler is the cleanest and the only way to change the event frequency.
ABT_sched_config config;
int new_freq = 16; // The default value is 50 (https://github.com/pmodels/argobots/blob/main/src/arch/abtd_env.c#L13)
ABT_sched_config_create(&config, ABT_sched_basic_freq, 16, ABT_sched_config_var_end);
1.2. Custom scheduler
You can call ABT_xstream_check_events() more frequently after calling ABT_info_trigger_print_all_thread_stacks() (e.g., when a global flag is on, a scheduler calls ABT_xstream_check_events() in every iteration).

2. ABT_info_trigger_print_all_thread_stacks()
ABT_info_trigger_print_all_thread_stacks() is designed for deadlock/livelock detection, so if your program is just (extremely) slow, ABT_info_trigger_print_all_thread_stacks() might not be a right routine to try.

> The first example I tried appeared to essentially defer dump until shutdown.
When one of your ULTs encounters a deadlock, the scheduling loop might not be called. You might want to set timeout for ABT_info_trigger_print_all_thread_stacks(). For example, the following test will forcibly print stacks after 3.0 seconds even if some execution streams have not reached ABT_xstream_check_events().
This is dangerous (I mean, it can dump a stack of a running ULT), so Argobots does not guarantee anything but it might be helpful to understand a deadlock issue sometimes.


3. Some tips
3.1. gdb
I would use gdb if it would be available to check a deadlock/performance issue. For example, if a program looks hanging, I will attach a debugger to that process and see what's happening.
3.2. libunwind for ABT_info_trigger_print_all_thread_stacks()
Unless you are an extremely skillful low-level programmer, I would recommend you enable libunwind for better understanding of stacks. By default, ABT_info_trigger_print_all_thread_stacks() dumps raw hex stack data.
3.3. "occasionally tied up in system calls"
I'm not sure if it's happening in the Argobots runtime (now Argobots uses futex for synchronization on external threads), but if you are calling ABT_info_trigger_print_all_thread_stacks() in a signal handler, please be aware that system calls terminate (e.g., futex, poll, or pthread_cond_wait) if a signal hits the process.
(Argobots synchronization implementation is aware of it and should not be affected by an external signal. This property is thoroughly tested: https://github.com/pmodels/argobots/blob/main/test/util/abttest.c#L245-L287)
Note that the user can call ABT_info_trigger_print_all_thread_stacks() on a normal thread without any problem. It is implemented just in an async-signal safe manner.
3.4. Stack dump
ABT_info_print_thread_stacks_in_pool() is a less invasive way to print stacks, especially if you know a list of pools. It prints stacks immediately. Basically, ABT_info_trigger_print_all_thread_stacks() sets a flag to call ABT_info_print_thread_stacks_in_pool() for all pools after all the execution streams stop in ABT_xstream_check_events().


From: Phil Carns via discuss <discuss at lists.argobots.org><mailto:discuss at lists.argobots.org>
Sent: Wednesday, April 14, 2021 2:18 PM
To: discuss at lists.argobots.org<mailto:discuss at lists.argobots.org> <discuss at lists.argobots.org><mailto:discuss at lists.argobots.org>
Cc: Carns, Philip H. <carns at mcs.anl.gov><mailto:carns at mcs.anl.gov>
Subject: [argobots-discuss] modifying scheduler event frequency?

Hi all,

Is there a clean way to change a scheduler's event frequency on the fly?

Browsing the API, I see two possibilities:

  *   set it when the scheduler is first created (using ABT_sched_basic_freq?)
  *   set it dynamically by manipulating the ABT_sched_get_data() pointer, but this seems especially dangerous since the sched data struct definition isn't public (i.e. it could cause memory corruption if the internal struct def changed)

For some context (in case there is a different way to go about this entirely), I'm trying to figure out how to get ABT_info_trigger_print_all_thread_stacks() to print information more quickly, which IIUC relies on getting the active schedulers to call get_events() sooner.

I'm happy to add some explicit ABT_thread_yield() shortly after the ABT_info_trigger_print_all_thread_stacks() to at least get the calling ES to execute it's scheduler loop immediately, but I think that won't matter much if it doesn't trip the frequency counter when I do it.

Without this (at least with the _wait scheduler and threads that are occasionally tied up in system calls) I think the stack dump is likely to trigger too late to display what I'm hoping to capture when I call it.  The first example I tried appeared to essentially defer dump until shutdown.


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.argobots.org/pipermail/discuss/attachments/20210414/013489e6/attachment.html>

More information about the discuss mailing list