Difference between revisions of "Tardis scheduling policy"

From CUC3
Jump to navigation Jump to search
import>Cen1001
import>Cen1001
Line 1: Line 1:
Now is a good time to review tardis's scheduling policy as we have had the machine in service for a few months. Whatever we decide upon needs to be given a fair chance to work, so after this consultation I am not intending to change the policy again until the Long Vac at the earliest. It's therefore important to have your say now.
+
We appear to need to take another look at tardis's scheduling policy. Greg and Catherine would like everyone to agree on a policy and then give it a reasonable trial before starting to tweak it again, so whatever we come up with we'd like to stick with until January 2009 at the earliest.
   
 
== Current policy ==
 
== Current policy ==
   
=== Fairshare and priority ===
+
=== Priority ===
   
  +
This determines what order Idle jobs are on the queue. Type 'showq' to see the sorted queue.
Every individual user has a fairshare target of 20% of the machine. If you go over that then you get penalties; under it and you get bonuses. There are also two QOS (quality of service) groups, which also have fairshare targets. These are based on how the machine was funded: the 'stuart' group (Stuart Althorpe's research group) get 52% and the 'portfolio' group (everyone else) get 48%.
 
   
 
Priority is currently: 20 * ( personal fairshare result + group fairshare result ) + job expansion factor .
Fairshare targets and usage can be seen by running 'diagnose -f'
 
   
 
Every individual user has a fairshare target of 20% of the machine. If you go over that usage then your fairshare result is negative; under it and it's positive. Everyone is also in one of four groups of users which have group fairshare targets. These are based on how the machine was funded: the 'stuart' group (Stuart Althorpe's research group) gets 38%, the 'michiel' group (Michiel Sprik) gets 20% (this includes their share of the Portfolio funding), the 'jochen' group (Jochem Blumberger) gets 14% and the 'portfolio' group (everyone else) gets 28%. Your group and your personal result are added, so other people in your group can make your fairshare negative by running lots of jobs.
The fairshare calculation takes the last six weeks of usage into account, decaying it at a rate of 0.8 per week.
 
   
  +
Fairshare targets and usage can be seen by running 'diagnose -f'. The research group targets as labelled as QAS targets rather than group targets- this is a technical consequence of the way we set the system up. 'group' to Maui means 'Unix group' not 'research group'.
Priority is currently: 20 * ( personal fairshare bonus/penalty + group fairshare bonus/penalty ) + job expansion factor .
 
  +
  +
The fairshare calculation works on 48-hour slots (the longest queue is 48 hours) and uses the last ten. Older slots count for less than recent ones.
   
 
Job expansion factor rises with time spent on the queue, but rises faster for short jobs. The reason for using that and not basic queue time is that it helps the very short (30 min) test jobs to run. It makes practically no difference when compared to the fairshare numbers, but ensures that every job eventually runs.
 
Job expansion factor rises with time spent on the queue, but rises faster for short jobs. The reason for using that and not basic queue time is that it helps the very short (30 min) test jobs to run. It makes practically no difference when compared to the fairshare numbers, but ensures that every job eventually runs.
Line 20: Line 22:
   
 
There is one throttling policy in use: any user may only have four jobs in the 'Idle' state at any given time. This avoids queue stuffing. However it does not help when one person has a very big fairshare bonus and submits a lot of jobs, because every job that gets to run is replaced in the queue immediately.
 
There is one throttling policy in use: any user may only have four jobs in the 'Idle' state at any given time. This avoids queue stuffing. However it does not help when one person has a very big fairshare bonus and submits a lot of jobs, because every job that gets to run is replaced in the queue immediately.
 
   
 
=== Reservation and Backfill ===
 
=== Reservation and Backfill ===
   
This policy isn't making much difference at the moment; I don't think it needs changing but I mention it for completeness. We make a reservation for the top queued job and then backfill other jobs around that (ie let them jump the queue if and only if they will not have any effect on the start time of the top queued job). This stops big parallel jobs being crowded out by small jobs, but only once they have got to the top of the queue. Without this the 32 proc jobs would almost never run.
+
The system makes a reservation for the top queued job and then backfills other jobs around that (ie lets them jump the queue if and only if they will not have any effect on the start time of the top queued job). This stops big parallel jobs being crowded out by small jobs, but only once they have got to the top of the queue. Without this the 64 proc jobs would almost never run.
   
=== The current problem ===
 
   
  +
=== Node groups ===
The Althorpe group currently have only user actively submitting jobs and little historical usage, so that user account gets a huge fairshare bonus and can crowd everyone else out of the machine. This is taking considerable time to correct because the fairshare memory is long and the deficit is large.
 
   
  +
We force multinode jobs run on a set of nodes that are all attached to the same Infiniband switch. This can cause very large jobs (64 and 32-way) to take a long time to start because the system has to clear most of a switch for them. Other jobs may appear to jump the queue in this situation but if you check then those jobs will be being started on switches other than the one the system has picked for the top job. The system picks the optimum switch for the job, ie the one where the job can start soonest if all the other jobs in the system run to their maximum walltime. The calculation is redone each time the system state changes.
=== Things we could change and their likely effects ===
 
   
 
== The current problem ==
Maui is amazingly configurable; any policy you can come up with we can probably find a way to make Maui do. Here are a few possibilities:
 
   
  +
32 and 64-way jobs are taking a very long time to start even with the above system.
* Shorter or fewer fairshare windows, so machine has shorter memory. We could reduce the windows to 48 hours which would mean the machine would only remember just under a fortnight of usage.
 
* Dilute group fairshare (ie give personal fairshare a bigger multiplier than 20)
 
* Drop group fairshare and possibly give Stuart's group bigger personal fairshare instead. The problem with this is working out how much is fair, as it would vary with the number of users in each category.
 
* Max processors per person limit. This would have to be quite high otherwise it reduces utilization.
 
* Max outstanding processor-seconds per person limit. Works well on machines with very variable job lengths. But again how much?
 
   
  +
I think the problem is this: one of these jobs reaches the top of the queue and the system books it a reservation and starts clearing space for it, but then a small, higher-priority job comes along and the top job loses its reservation. The small job takes some of the space. The system can't start the large job so it backfills the rest of the queue into what space is left. Once the small job starts the large job is top but has to start clearing space to run in all over again...and then another small job comes along.
--[[User:cen1001|Catherine]] 16:01, 16 March 2007 (GMT)
 
   
  +
This is made worse by the tendency of groups which run small jobs to be under their fairshare target and the large job groups to be over.
   
 
=== Things we could change and their likely effects ===
My suggestion is that we could increase the priority penalty for each jobs that is already running. In other words
 
if I have already 10 jobs running, my jobs in queue should get a lower priority. In this way if the machine is empty
 
I can still run many jobs, which will not be possible if a "Max processors per person limit" or "Max outstanding processor-seconds per person limit" is introduced.
 
 
 
--[[User:ic247|Ivan]] 10:47, 17 March 2007 (GMT)
 
 
My suggestion would be to limit the maximum processors per person (set high) if there are jobs from other users queuing, but allow one user to use the whole machine if it's empty. I also think a shorter fairshare window is needed. Usage of tardis has fluctuated a lot recently so I think it should reflect this.
 
 
--[[User:cja49|Chris]] 14:20, 19 March 2007 (GMT)
 
 
 
I think Ivan's suggestion is an excellent solution. A max processors/max processor-seconds approach could stop someone from using the full resources available if tardis is empty, yet wouldn't penalise queued jobs submitted by people who are occupying a large amount of the cluster.
 
 
A 6 week memory might be on the long side: I go through periods of running just short test jobs (mainly on my workstation) and periods when I'm running large numbers of fairly demanding calculations. I get these through very quickly because of my large fairshare factor (which is good for me, but maybe less good for those who have to wait whilst I occupy a huge chunk of a cluster for several days).
 
 
--[[User:jss43|james]] 17:05, 19 March 2007 (GMT)
 
 
A MAXPS or MAXPROC limit causes excess jobs to be blocked, so it ''does'' penalise queued jobs submitted by people who are occupying large parts of the cluster already, by not allowing them to run at all. We are using those policies on destiny and mek-quake successfully at the moment.
 
 
Ivan's suggestion seems to me to be equivalent to fairshare with a very short memory. This would work but I think setting the memory too short would counterproductive.
 
 
Chris's suggestion of a variable MAXPROC limit is an interesting one; Maui can do something like this but I'm not sure of the details. I will look them up. The thing that approach suffers from is that you still get the latency of waiting for a slot after one person has filled up the entire machine, even if their queued jobs are then blocked.
 
 
--[[User:cen1001|Catherine]] 09:03, 20 March 2007 (GMT)
 
 
Yes, I think that 6 weeks' memory is probably a bit too long: I would suggest we try 3 weeks and see how this works out.
 
 
Also, another suggestion: it would be useful if the script could be configured to send out a weekly or monthly automated summary of usage. This would help identify any problems that needed rectifying in the longer term.
 
 
--[[User:sca10|stuart]] 11:03, 20 March 2007 (GMT)
 
   
The policy was eventually changed to this:
 
   
  +
* Increase the reservation depth to 2 or 3, causing the system to make reservations for the top 2 or 3 queued jobs, not just the top one. This will reduce overall utilization but on average help larger jobs because they won't keep losing their reservations. Small jobs tend to run fast enough that reservations are not an issue for them.
Fairshare is now calculated over 10 48-hour periods decaying with a factor 0.8, reducing the total fairshare memory and making recent usage even more highly weighted than older usage.
 
   
  +
* Give a straight priority boost to large jobs to try to keep them at the top of the queue. More complex to configure. Would need to decide how much is enough and define a 'large job'.
--[[User:cen1001|Catherine]] 16:32, 19 April 2007 (BST)
 

Revision as of 17:30, 9 September 2008

We appear to need to take another look at tardis's scheduling policy. Greg and Catherine would like everyone to agree on a policy and then give it a reasonable trial before starting to tweak it again, so whatever we come up with we'd like to stick with until January 2009 at the earliest.

Current policy

Priority

This determines what order Idle jobs are on the queue. Type 'showq' to see the sorted queue.

Priority is currently: 20 * ( personal fairshare result + group fairshare result ) + job expansion factor .

Every individual user has a fairshare target of 20% of the machine. If you go over that usage then your fairshare result is negative; under it and it's positive. Everyone is also in one of four groups of users which have group fairshare targets. These are based on how the machine was funded: the 'stuart' group (Stuart Althorpe's research group) gets 38%, the 'michiel' group (Michiel Sprik) gets 20% (this includes their share of the Portfolio funding), the 'jochen' group (Jochem Blumberger) gets 14% and the 'portfolio' group (everyone else) gets 28%. Your group and your personal result are added, so other people in your group can make your fairshare negative by running lots of jobs.

Fairshare targets and usage can be seen by running 'diagnose -f'. The research group targets as labelled as QAS targets rather than group targets- this is a technical consequence of the way we set the system up. 'group' to Maui means 'Unix group' not 'research group'.

The fairshare calculation works on 48-hour slots (the longest queue is 48 hours) and uses the last ten. Older slots count for less than recent ones.

Job expansion factor rises with time spent on the queue, but rises faster for short jobs. The reason for using that and not basic queue time is that it helps the very short (30 min) test jobs to run. It makes practically no difference when compared to the fairshare numbers, but ensures that every job eventually runs.

Priority calculations for all queued jobs can be seen by running 'diagnose -p'

Throttling policies

There is one throttling policy in use: any user may only have four jobs in the 'Idle' state at any given time. This avoids queue stuffing. However it does not help when one person has a very big fairshare bonus and submits a lot of jobs, because every job that gets to run is replaced in the queue immediately.

Reservation and Backfill

The system makes a reservation for the top queued job and then backfills other jobs around that (ie lets them jump the queue if and only if they will not have any effect on the start time of the top queued job). This stops big parallel jobs being crowded out by small jobs, but only once they have got to the top of the queue. Without this the 64 proc jobs would almost never run.


Node groups

We force multinode jobs run on a set of nodes that are all attached to the same Infiniband switch. This can cause very large jobs (64 and 32-way) to take a long time to start because the system has to clear most of a switch for them. Other jobs may appear to jump the queue in this situation but if you check then those jobs will be being started on switches other than the one the system has picked for the top job. The system picks the optimum switch for the job, ie the one where the job can start soonest if all the other jobs in the system run to their maximum walltime. The calculation is redone each time the system state changes.

The current problem

32 and 64-way jobs are taking a very long time to start even with the above system.

I think the problem is this: one of these jobs reaches the top of the queue and the system books it a reservation and starts clearing space for it, but then a small, higher-priority job comes along and the top job loses its reservation. The small job takes some of the space. The system can't start the large job so it backfills the rest of the queue into what space is left. Once the small job starts the large job is top but has to start clearing space to run in all over again...and then another small job comes along.

This is made worse by the tendency of groups which run small jobs to be under their fairshare target and the large job groups to be over.

Things we could change and their likely effects

  • Increase the reservation depth to 2 or 3, causing the system to make reservations for the top 2 or 3 queued jobs, not just the top one. This will reduce overall utilization but on average help larger jobs because they won't keep losing their reservations. Small jobs tend to run fast enough that reservations are not an issue for them.
  • Give a straight priority boost to large jobs to try to keep them at the top of the queue. More complex to configure. Would need to decide how much is enough and define a 'large job'.