Multi-clustering can be defined as the sharing of compute jobs between two or more independently controlled clustergrids. This could be desirable, for example, if the owners of the clusters realize that there are times when one cluster is lightly-loaded while another has a large pending list. In this case, it makes sense to make use of the idle resources, at least until the "local" demand for this cluster rises again.
We define the "local" cluster as the cluster with which a user would normally interact (qsub, qstat, etc), and "remote" cluster as the one to which jobs are dispatched (if deemed appropriate).
One way to implement multi-clustering is with transfer queues. The idea is that jobs dispatched to a transfer queue get submitted to a remote cluster; this would happen transparently to the user. To have jobs going in both directions, simply use transfer queues on both sides.
The prototype described here implements the following ideas:
a load sensor which gives the number of pending jobs in the remote cluster
a starter method which takes the job script and submits it to the remote cluster
suspend, resume, and terminate methods which act upon jobs sent to the remote cluster
a transfer queue which implements the methods described above, and which has a load threshold based upon the number of pending jobs in the remote cluster.
The scripts used to implement the prototype can be downloaded here:
both clusters share a common "namespace" or "administrative domain". By this we mean: common usernames/UIDs/GIDs, common filesystem(s), mutually accessible hosts and hostnames, behind the same firewalls, etc.
resource requests are simply passed along without modification. This could lead to errors if a particular resource is not defined in the remote cluster.
once a job is submitted to the remote cluster, it will remain in the pending list for that cluster until it either runs or is manually deleted. If meanwhile some local resources are freed, the jobs which are pending at the remote cluster will not get rerouted back to the local cluster.
To implement this prototype, create a new queue (the transfer queue) on any existing host. Configure it as shown below. Implement the load sensor on any host (it would make sense to implement it on the same host where the transfer queue lives, but that's not necessary). Configure that host as shown below.
NOTE: it is suggested that the transfer queue be implemented on the local master host, to keep things simple and to reduce communication latency.
NOTE: make sure this host is both a submit and admin host on the remote cluster.
NOTE: please modify the scripts to match your local environment. See the scripts for site-specific variables.
Test it by submitting jobs directly to the transfer queue, eg,
qsub -q transfer1 -o output.log -j y sleeper.sh
In practice, you would not submit jobs directly to this special queue. Rather, one of the following approaches would be taken:
if all jobs are "generic" enough, then it might not matter if they run in the local or remote cluster. In this case, jobs would just be submitted normally, and those which happen to end up in the transfer queue will run remotely
if certain jobs cannot be run on the remote cluster for whatever reason (licensing restrictions, data/binaries inaccessible remotely), then a user-defined complex with a boolean resource (eg, "remote") should be attached to the transfer queue(s). Jobs which cannot be run remotely could be submitted with a resource request "-l remote=false". This will prevent those jobs from being dispatched to the transfer queue(s), and hence will not be dispatched remotely.
sr3-umpk-01$ qconf -sq transfer1 qname transfer1 hostname tcf-b060 load_thresholds mpk27jobs=5 qtype BATCH slots 40 prolog NONE epilog NONE starter_method /sge/TransferQueues/transfer_starter.sh suspend_method /sge/TransferQueues/transfer_suspend.sh resume_method /sge/TransferQueues/transfer_resume.sh terminate_method /sge/TransferQueues/transfer_terminate.sh
sr3-umpk-01$ qconf -sconf tcf-b060 tcf-b060: load_sensor /sge/TransferQueues/clusterload.sh load_report_time 00:00:09
the total number of "foreign" jobs in a system (both running and pending) can be specified by setting the number of slots in the transfer queue. It is recommended that this number be set to some fraction of the number of remote slots (eg, 1/3 or 1/4), for a couple of reasons:
if many jobs are dispatched to the remote cluster and are pending there, and the remote cluster gets flooded with higher priority jobs, then all those pending jobs could get trapped, instead of potentially running on free local resources
if there are too many jobs dispatched to the remote cluster, then the polling mechanism for the transfer queue becomes too much of a load (but see improvements below)
jobs only get sent to the remote cluster if it is "idle". The definition of idle is having N or fewer jobs in the remote pending list; N is configurable but might be something like 10% of the total number of remote job slots. This number is used in the load threshold for the transfer queue.
if using Grid Engine Enterprise Edition, "foreign" jobs can be assigned a particular project category, either voluntarily in the job request or by giving it an assigned category, and the share of this project in the local cluster modified appropriately. The remote project is set in the transfer starter method.
see comments in the scripts to see how certain functionality was implemented and why other choices were not used.
if this setup is to be used for a high-throughput environment (large numbers of relatively short jobs, especially array jobs), the load sensor will not be able to provide a very accurate measure of the size of the pending list, and might unnecessarily prevent jobs from being dispatched. In this case, you might consider disabling the load threshold, and instead setting the number of slots to a smaller number (eg, 1/10 the number of remote slots). This will meter out the small jobs in a steady but limited stream, thus avoiding the possibility of dispatching too many jobs and overloading the cluster. It would also be useful here to set the polling time to something lower, like 5 seconds. CAUTION: do not set the polling time to anything less than 5 seconds, otherwise the remote cluster will become overloaded with status requests.
One convenient way to do resource mapping, especially if the remote cluster has systems of widely varying capabilities, is to use multiple transfer queues for the different types of systems (eg, different Oses, different memory sizes, etc). By setting the resources on the local transfer queue to match a specific type of system on the remote cluster, you can ensure that jobs get dispatched only when the correct type of system is available.
There are some improvements which could be made to the current implementation of methods:
reduced polling: the currently implementation uses
calls for gathering required information on both the local and
remote clusters. This can lead to excessive overhead. Two ways in
which this can be improved are:
eliminate the initial
qstat in the starter
method, and instead obtain job submission information directly from
the job spool directory ($SGE_JOB_SPOOL_DIR/config). One limitation
is that this currently will not provide information about the hard
and soft requests for the job. This might be acceptable if, for
example, you restrict the transfer queue to accept only jobs with
certain criteria. In this case, you will know the resource
requirements for the job and can hard code them. Multiple transfer
queues could be used to implement multiple criteria sets.
instead of polling the remote cluster for each job individually, a separate polling process could be used to get a list of all pending/running/suspended jobs from the remote cluster. Then, you would only need to check that list, rather than do multiple repeated qstats.
invalid resource requests: if a request is invalid, have an error handler which sends the job back to the local cluster and marks it ineligible to run at the remote cluster. Currently, the job is just killed.
resource request mapping: if special resources are requested, have a way to map them to equivalent resources at the remote cluster.
accounting: have a sensible way to retrieve accounting
information from the remote cluster. Perhaps, have a '
command which will get the remote JID from the local accounting
record and then do a remote regular
also have a way to map
commands to the remote cluster. Again, have a
version of these commands.
The implementation described here is limited in terms of transfering jobs between administrative domains (eg, different file/user namespaces, different firewalls,etc). In order to extend the concept to multiple domains, the following extensions to the MultiCluster concept would require putting additional functionality into the Grid Engine code, and most likely, integration with functionality provided by other software:
Cross domain security and identity.
Firewalls, NATs, etc.
Data transfer between domains.
Brokerage for intelligent routing of jobs based on least cost, minimal time, best resource, or similar criteria.
Sharing of resources across domains, e.g. software licenses; policies related to such sharing models.
Cross domain accounting, billing, monitoring.
Co-scheduling across domains for distributed apps (e.g. computation and visualisation).
Cross domain error tracking.
Utilization of standardized interfaces for global grids (e.g. OGSA).
In addition, cross-domain issues tend to rely more upon standard APIs and protocols. Currently, these are not well-defined; much work towards defining these standards is occuring in standards bodies such as the Global Grid Forum (GGF).