gang_sched(7)

NAME

gang_sched — Gang Scheduler

DESCRIPTION

The gang scheduler permits a set of MPI (Message Passing Interface) processes, or multiple threads from a single process, to be scheduled concurrently as a group.

Gang scheduling is enabled and disabled by setting the MP_GANG environment variable to ON or OFF.

The gang scheduling feature can significantly improve parallel application performance in loaded timeshare environments that are oversubscribed. Oversubscription occurs when the total number of runnable parallel threads, runnable MPI processes, and other runnable processes exceeds the number of processors in the system.

Gang scheduling also permits low-latency interactions among threads in shared-memory parallel applications.

Only applications using the HP-UX V11.0 MPI or pthread libraries can be gang scheduled. Because HP compiler parallelism is primarily built on the pthread library, programs compiled with HP compilers can benefit from gang scheduling.

INTERFACE

The HP-UX gang scheduler is enabled and disabled using an environment variable. The variable is defined as:

MP_GANG [ON | OFF]

Setting MP_GANG to ON enables gang scheduling and setting it to OFF disables it. If MP_GANG is not set, or if it is set to an undefined value, no action is taken.

Gang scheduling is a process attribute that is inherited by child processes created by fork (see fork(2)). The state of gang scheduling for a process can change only following a call to exec (see exec(2)).

BEHAVIOR

After the MP_GANG environment variable is set to ON, any MPI or pthread application to execute and find this variable will enable gang scheduling for that process.

Only the pthread and MPI libraries query the MP_GANG variable--the operating system does not.

Gang scheduling is an inherited process attribute. When a process with gang scheduling enabled creates a child process, the following occurs:

The child process inherits the gang scheduling attribute.
A new gang is formed for the child process. The child does not become part of its parent's gang.

The gang scheduler is engaged only when a gang consists of multiple threads. For a pthread application, this is when a second thread is created. For an MPI application, it is when a second process is added.

As a process creates threads, the new threads are added to the process's gang if gang scheduling is enabled for the process. However, once the size of a gang equals the number of processors in the system, the following occurs:

New threads or processes are not added to the gang.
The gang remains intact and continues to be gang scheduled.
The spill-over threads are scheduled with the regular timeshare policies.
If threads in the gang exit (thus making room available), the spill-over threads are not added into the gang. However, newly created threads are added into the gang when room is available.

MPI processes are allocated statically at the beginning of execution. When MP_GANG is set to ON, all processes in an MPI application are made part of the same gang.

Thread and process priorities for gangs are managed identically to timeshare policy. The timeshare priority scheduler determines when to schedule a gang and adheres to the timeshare policies.

Although it is likely that scheduling a gang will preempt one or more higher priority timeshare threads, over the long run the gang scheduler policy is generally fair. All threads in a gang will have been highest priority by the time a gang is scheduled. Because all threads in a gang must execute concurrently, some threads do not execute when they are highest priority (the threads must wait until all other threads have also been selected, allowing other processes to run first).

Gangs are scheduled for a single time-slice. The time-slice is the same for all threads in the system, whether gang-scheduled or not.

When a single gang executes on a system, the gang's threads are assigned to processors in the system and are not migrated to different processors.

In an oversubscribed system with multiple gangs, all gangs are periodically moved in order to give an equalized percentage of CPU time to each of the different threads. This rebalancing occurs every few seconds.

EXTERNAL INFLUENCES

Environment Variables

The following environment variables affect gang scheduling of processes:

MP_GANG enables (when set to ON) and disables (when set to OFF) gang scheduling of processes. For details see the INTERFACE section of this man page.
MP_NUMBER_OF_THREADS specifies the number of processors available to execute programs compiled for parallel execution. If not set, the default is the number of processors in the system.

PERFORMANCE

Gang scheduling ensures that all runnable threads and processes in a gang are scheduled simultaneously. This improves the synchronization latency in parallel applications. For instance, threads waiting at a barrier do not have to wait for currently unscheduled threads.

However, applications with lengthy parallel regions and infrequent synchronization may perform best when not gang scheduled. For those applications, some threads can be scheduled even if all threads are not scheduled at once.

A gang-scheduled application's performance can be affected by the number of gang-scheduled applications on a system, and by the number of threads in each. The gang scheduler assigns parallel applications to CPUs using a "best fit" algorithm that attempts to minimize CPU overlap among applications.

On systems with complex workloads including gangs of varying sizes, or odd combinations of sizes, the workload may not optimally match the number of CPUs available. In this situation an application may perform better when not gang scheduled, thus enabling some threads to be scheduled rather than waiting for all threads to be scheduled as a gang.

Scheduling Overhead

Gang scheduling incurs overhead when the scheduler collects a set of threads, assigns a set of processors to the threads, and rendezvous the set of threads and processors to achieve concurrent execution.

On an idle system, the gang scheduling overhead can be seen in the execution time of a single parallel application.

Kernel Blocking of Threads

If a thread from a gang blocks in the kernel, the thread's processor is available to run other non-gang-scheduled threads. When the blocked thread resumes and its gang is currently running, the thread can join the other ganged threads without having to rendezvous again.

In a multi-gang environment, thread blocking can result in lower throughput. This occurs if an application's threads block often in the kernel for long periods of time.

Preempting by Realtime Threads

Gang-scheduled threads can be preempted from execution by realtime threads. This affects only the gang-scheduled thread running on the processor being preempted by a realtime thread. The remaining threads of the gang continue to run through the end of their time-slice.

RESTRICTIONS

For this implementation of gang scheduling, the following restrictions exist. Some of these may be removed in future releases.

Gang scheduling of processes being debugged is not supported. When a debugger attaches to a process, gang scheduling for the process is disabled. This avoids gang scheduling processes with one or more threads stopped by a debugger.
Gang scheduling is completely shut down when Process Resource Manager (PRM) is enabled.
If a gang-scheduled process is selected to be swapped out, the process will not be gang-scheduled when it is swapped back in.
Realtime processes are not gang-scheduled.
Gang scheduling is only supported for processes with timeshare scheduling policies.
When a gang-scheduled process contains the maximum number of threads (or the maximum number of processes, for MPI applications), threads or processes created after this point are not scheduled as part of the gang. For details see the BEHAVIOR section of this man page.
Multiprocess applications that do not use MPI are not supported by the gang scheduler.
Gang scheduling is not supported for PTHREAD_SCOPE_PROCESS threads. From release 11i Version 1.6 of HP-UX, the default scheduling contention scope for threads is PTHREAD_SCOPE_PROCESS. If any PTHREAD_SCOPE_PROCESS threads are created by an application, the initial thread will be treated as a PTHREAD_SCOPE_PROCESS.

FILES

The following are libraries used in providing gang scheduling:

/usr/lib/libpthread.1: The pthread library.
/opt/mpi: The directory containing MPI libraries and MPI software. HP MPI is an optional product.