Re: QoS scheduler

[Vitor Curado <curado.vitor@gmail.com>]

Indeed I didn't specify what my project is about... My goal is to
benchmark several QoS process schedulers, comparing them to the native
kernel scheduler. I didn't, however, decided how will the benchmarking
be done. Any sugestions?


On 7/29/05, Stephen Pollei <stephen.pollei@gmail.com> wrote:
> On 7/29/05, Vitor Curado <curado.vitor@gmail.com> wrote:
> > You assumed right, Stephen: I'm interested in QoS process scheduling,
> > sorry for not specifying it...
> >
> > I'm taking a deeper look at the qlinux, ckrm and the plugsched
> > schedulers, if you have any more links, please send them to me...
> Also you didn't specify what kind of clustering you are doing and for
> what ultimate purpose.
> 
> http://www.beowulf.org/
> http://www-unix.mcs.anl.gov/mpi/implementations.html
> http://www.csm.ornl.gov/pvm/pvm_home.html
> http://www.open-mpi.org/
> 
> http://openmosix.sourceforge.net/
> http://www.mosix.org/
> 
> http://www.remote-dba.cc/teas_aegis_rac06.htm
> http://www.dba-oracle.com/bp/bp_book1_rac.htm
> Oracle DB Real Application Clusters (RAC)
> transparent application failover (TAF)
> 
> http://pgcluster.projects.postgresql.org/feature.html
> http://dev.mysql.com/doc/mysql/en/replication.html
> 
> High Availability (HA)
> High Performance Computing (HPC)
> 
> That can strongly effect what solutions you would want to look at.
> For instance if you were running a render farm, or a scientific
> compute beowulf cluster, then
> your "scheduling" will be handled more in the MPI or PVM code perhaps.
> The running processes themselves would most likely be using something
> like SCHED_BATCH, with larger than usual time-slices. Maybe you
> monitor how many mips actually get consumed and then adjust which
> nodes get scheduled with what, or how many work units get handed out
> to get back to fairness.
> 
> clock_t times(struct tms *buf);
> int getrusage(int who, struct rusage *usage);
> to track system and user time is about on track, but I think someone
> might be able to fool you, if thats all you could use to account for
> cpu time taken from another userland process.
> 
> So maybe you just need better reporting/accounting hooks and then you
> can do the rest in userland?
> 
> > On 7/28/05, Wes Felter <wesley@felter.org> wrote:
> > > Vitor Curado wrote:
> > > > I'm working on a research about QoS schedulers for Linux clusters.
> > > > Moreover, the ideal would be that the scheduler is implemented
> > > > altering the native kernel scheduler. I'm kind of having trouble to
> > > > find such schedulers, can anybody help me out?
> > >
> > > http://lass.cs.umass.edu/software/qlinux/
> > > http://ckrm.sourceforge.net/
> 
> That qlinux one is new to me. I notice that the 2.6 kernel has support
> for modular plugable disk I/O and network schedulers now.
> So  a Hierarchical Start Time Fair Queuing (H-SFQ) network packet
> scheduler module could be made.
> 
> I wonder how that Cello scheduler would stack-up to AS, Deadline, cfq,
> noop, etc etc.
> 
> The qlinux cpu scheduler would be best to use plugsched for use with 2.6.x
> 
> --
> http://dmoz.org/profiles/pollei.html
> http://sourceforge.net/users/stephen_pollei/
> http://www.orkut.com/Profile.aspx?uid=2455954990164098214
> http://stephen_pollei.home.comcast.net/
>