Differences

This shows you the differences between two versions of the page.

--- wiki:disk_reservation [2016/11/29 23:01] – [Server prologue/epilogue] neyron
+++ wiki:disk_reservation [2016/11/30 10:26] – neyron
@@ Line 2: / Line 2: @@
 E.g. Each node has "cpus" and "disks", we want to allow users to create jobs to reserve disks for a long period, and eventually request smaller compute jobs on resources on which those disks are reserved.
 This would give users the capability of keeping data on disk for a longer time than the compute time, e.g. a more persistent storage for data.
 ===== Setup =====
 We create //disk// resources, then setup some coupling so that a //compute// resource is **tagged** when a user has disks reserved on it. Tagging in done in the //disk// property of the //compute// resource (type //default//), with the syntax: ''/userid1:#disks/userid2:#disks/''.
@@ Line 156: / Line 157: @@
 Mind setting a relevant priority for the new admission rule.
-Finally, just like for deploy jobs, we allow only whole nodes for //compute with disk// jobs. For that, we edit the corresponding rule: ''oaradmissionrules -m 9 -e''
+Finally, just like for deploy jobs, we might want to allow only whole nodes for //compute with disk// jobs. For that, we edit the corresponding rule: ''oaradmissionrules -m 9 -e''
 <code perl>
 -# Restrict allowed properties for deploy jobs to force requesting entire nodes
@@ Line 276: / Line 277: @@
   * This workaround does not allow to place and show (schedule) the //compute// jobs in the future (i.e. before the //disk// job is running)... While not scheduled, //compute// jobs can be submitted ahead of time however. Also the scheduling decision could place //compute// jobs after the end of the running //disk// job, so that they would actually not launch in fine.
   * Users may want to select specific disks/hosts when using only a part of the reserved disks for a given compute job. This mechanism does not allow it.
+  * The ''reserve_disk'' and ''with_disk'' job types do not allow to reserve at once (in a single ''oarsub'' command) both the disk and compute resources.
+  * A //with_disk// compute job could easily suffer from starvation since it will only be scheduled once the disk job is started:
+    * While batch jobs which are not started yet will be move with regard to previous scheduling decisions, some may have started before the disk property of the resources is changed, making resources whose disks are reserved unavailable for the duration of those jobs.
+    * Advance reservations could also be accepted on the resources: resources are //booked// upon submission acceptation for an advance reservation, possibly before the disk property is changed.