veyepar: Gang Scheduling in Linux Kernel Scheduler

Author(s):	Nikunj A Dadhania
Location	Studio
Date	jan Thu 19	Days Raw Files
Start	16:40	First Raw Start	error-in-template
Duration	0:50:00	Offset	None
End	17:30	Last Raw End
Chapters
Total cuts_time	None min.

http://lca2012.linux.org.au/schedule/171/view_talk
raw-playlist raw-mp4-playlist encoded-files-playlist mp4 svg png

assets release.pdf Gang_Scheduling_in_Linux_Kernel_Scheduler.json
logs
Admin: episode episode list cut list raw files day marks day marks day image_files

State:

Locked:

clear this to unlock

Locked by:

user/process that locked.

Start:

initially scheduled time from master, adjusted to match reality

Duration:

length in hh:mm:ss

Name:

Video Title (shows in video search results)

Emails:

email(s) of the presenter(s)

Released:

has someone authorised pubication

Normalise:

Channelcopy:

m=mono, 01=copy left to right, 10=right to left, 00=ignore.

Thumbnail:

filename.png

Description:

markdown

Operating System scheduler assigns the most eligible process to run on a CPU. On a multi-CPU system, tasks are assigned to CPUs independent of each other. The scheduler may not look into inter-process dependencies while assigning processes. Hence, at any given point in time, the set of processes that is running on the system might not be the optimal one. Gang Scheduling is a method to schedule a group of inter-dependent processes on different CPUs simultaneously. In this talk, we will discuss the design alternatives for Gang Scheduling, the experiments with KVM Hypervisor and the results.

In a multi-CPU Virtual Machine(VM), processes representing virtual CPUs(VCPUs) are not guaranteed to be scheduled simultaneously. This challenges the assumptions held by the VM about the CPUs on which it is running. In turn, this can inhibit the performance of processes running inside the VM in several ways.

For example, suppose one of the VCPUs takes a spinlock (defined within a VM) and gets preempted by the scheduler before it could release that spinlock. This forces other VCPUs wanting to acquire the same spinlock to spin for a long time wasting lot of CPU cycles. The problem is popularly known as Lock-Holder Preemption. This results in higher synchronization latency for workloads running within a VM. These latencies affect the workloads in an adverse manner. In contrast, with Gang Scheduling, all the VCPUs are scheduled simultaneously on a group of CPUs. These changes help to reduce synchronization latencies, resulting in workloads performing better in some cases.

Experiments were carried out to measure the benefits of gang scheduling with various workloads (CPU bound, IO bound, CPU+IO bound) in the KVM Hypervisor. In some cases, improvements more than 50% were observed.

Comment:

production notes