Setting up the Fair Share Scheduler (FSS) on Solaris

One of the classic problems in multi-user operating systems is that, by default, the scheduler will treat all processes or threads equally. Priorities were in part introduced to provide some way of differentiating between tasks, but they do nothing to make sure there's a fair distribution of CPU time across all users. In other words, the scheduler assumes all users and their processes are equal and should be treated the same.

Unix and Linux based operating systems then began to introduce abstractions to group or pool threads together according to some user specified notion of task. There's usually some amount of granularity involved here, but to keep this short, we'll work with "projects".

In Solaris (and its open source derivatives), a project is used to group threads together and can be tied to user attributes and/or various resource management features (there are many). Here's one way of setting up a shared system so that all users get the same amount of CPU time.

Add a project for each user with projadd(1). The '-p 123' argument is a project id, I suggest just using the user id for that user. The name of the project "user.bob" must begin with "user." so the system will automatically assign that user's processes to this project.

# projadd -p 123 -c "bobs project" -U bob -K "project.cpu-shares=(privileged,5,none)" user.bob

Which will look like the last line in /etc/project:

$ cat /etc/project
system:0::::
user.root:1::::
noproject:2::::
default:3::::
group.staff:10::::
user.bob:123:bobs project:bob::project.cpu-shares=(privileged,5,none)

The number of shares is not a percentage or an absolute value. It's a proportionate amount in relation to other users' shares: projects with more shares will get more than others. The values themselves are only significant when compared to other projects.

Then make FSS the default scheduling class:

# dispadmin -D FSS

The per user projects will only start to be used for new user logins, so you'll need to ask your users to logout and back in. After that, you can monitor per project stats with 'prstat -J':

$ prstat -J
   PID USERNAME  SIZE   RSS STATE  PRI NICE      TIME  CPU PROCESS/NLWP
   422 root        0K    0K sleep   99  -20  15:40:29 0.6% zpool-build/85
   518 root       10M 4384K sleep   29    0   1:51:33 0.0% automountd/8
  1212 bob        12M 9156K cpu4    49    0   0:00:00 0.0% gcc/1
 PROJID   NPROC  SWAP   RSS MEMORY      TIME  CPU PROJECT
   123      10   145M   76M   0.2%   0:00:04 1.2% user.bob
     0      61   183M  163M   0.3%  20:25:34 0.6% system
     3     274  1087M 1050M   2.1%   2:16:54 0.1% default