Hi; You're right that I wasn't thinking about the most general cases that can occur in workflow and parallel/distributed programs, such as MPI or UPC programs. In particular, I was only thinking about things that can be declaratively described and don't require the provision of code that needs to run "inside" the scheduling infrastructure. Let's consider workflow and job dependencies first. Static workflows and job dependencies can be declaratively described in the form of XML infosets using standardized terminology. For these I claim that the extension mechanisms I've described are sufficient. That is, supporting static workflows and job dependencies is a matter of defining the appropriate standardized description syntax and semantics and supporting extended versions of such is a matter of agreeing on how extensions to the descriptions can be made (which are covered by the mechanisms I've already included). Dynamic workflows and job dependencies require that a client supply application-specific code that can be run inside the scheduling infrastructure in order to supply dynamically computed decisions. This requires an additional extension mechanism beyond the ones that I listed. Note that whether the client describes the decision making code in terms of something like BPEL or in terms of something like a Java servlet that the scheduling infrastructure runs is a second-order issue. In both cases the client is supplying code that gets run inside the scheduling infrastructure. So, you are right that we need the ability to run client-supplied code inside the scheduling infrastructure for some of the extensions we might contemplate. That said, I would argue that we should save these kinds of extensions for later in our deliberations since they are far more complicated to get right than the ones I listed. But we should definitely keep them in mind. Regarding MPI/UPC and other forms of parallel/distributed programs (e.g. PVM): There is a declarative aspect that is visible to clients and an internal, implementation aspect that I would argue should not be visible in the interface between clients and schedulers. Let's consider an MPI program that is based on the MPICH infrastructure consisting of SMPD daemons running on each compute node used by the program. A client will specify the MPI program to run, which MPI infrastructure it expects, and the relevant MPI-related arguments to supply (as well as other arguments, environment variables, etc.). This can all be described and encoded as an XML infoset. The scheduling infrastructure will internally need to implement the MPICH SMPD daemon-based infrastructure, but the details of that aren't visible in the job scheduling interface. An interesting question is whether MPICH's implementation aspects need to become visible when we consider scheduler-scheduler interactions. If an MPI program can span multiple clusters then the relevant SMPD daemons from multiple clusters need to be put in touch with each other. In the case of MPICH, I believe the main thing needed is that the root SMPD daemon receive a list of the IP addresses of all the compute nodes that will participate in a given MPI program. In that case, the server-server aspects of scheduling an MPI program mainly have to do with allocating the appropriate number of compute nodes - and getting their names - from appropriate compute clusters that have indicated that they support the MPICH SMPD infrastructure. So I hypothesize that support for parallel/distributed programs is mainly a matter of defining the appropriate declarative standards and doesn't require any additional extension mechanisms beyond those I've already described. I would, of course, be very interested to learn of examples where this is not enough. Marvin. ________________________________ From: Balle, Susanne [mailto:Susanne.Balle@hp.com] Sent: Monday, May 01, 2006 7:42 AM To: Marvin Theimer Cc: ogsa-wg@ggf.org Subject: RE: [ogsa-wg] Thoughts on extensions mechanisms for the HPC profile work Marvin I think this is a good start. I did find some areas missing such as Workflow and support for Job Dependencies as well as extensions for MPI/UPC programs. I like the "object-oriented" approach and agree with Dave that being able to specify more complex expressions is important and is in my opinion a requirement for ease of use. If you launch a 1000 job you want to be able to do query on groups of jobs without having to specify the individual jobs. Regards Susanne -----Original Message----- From: owner-ogsa-wg@ggf.org [mailto:owner-ogsa-wg@ggf.org] On Behalf Of Marvin Theimer Sent: Friday, April 28, 2006 10:06 PM To: ogsa-wg@ggf.org Subject: [ogsa-wg] Thoughts on extensions mechanisms for the HPC profile work Hi; This email is intended to describe my views of the set of extension mechanisms that are both necessary and sufficient to implement the common cases that we have identified for the HPC profile work (see the document "HPC Use Cases - Base Case and Common Cases", a preliminary draft of which I sent out to the ogsa-wg mailing list several weeks ago). These views are in large part derived from ongoing discussions that Chris Smith and I have been having about the subject of interoperable job scheduling designs. This email is intended to start a discussion about extension mechanisms rather than define the "answer" to this topic. So please do reply with suggestions for any changes and extensions (:-)) you feel are needed. Marvin. Additive vs. modifying extensions At a high level, there are two types of extensions that one might consider: 1. * Purely additive extensions. 2. * Extensions that modify the semantics of the underlying base-level design. Purely additive extensions that, for example, add strictly new functionality to an interface or that define additional resource types that clients and schedulers can refer to, seem fairly straight-forward to support. Modifying extensions fall into two categories: 3. * Base case semantics remain unchanged to parties operating at the base (i.e. un-extended) level. 4. * Base case semantics change for parties operating at the base level. Modifying extensions that leave the base-level semantics unchanged are straight-forward to incorporate. An example is adding at-most once semantics to interface requests. These operations now have more tightly defined failure semantics, but their functional semantics remain unchanged and base-level clients can safely ignore the extended semantics. Extensions that change base-level semantics should be disallowed since they violate the fundamental premise of base-level interoperability. An example of such an extension would be having the creation of jobs at a particular (extended) scheduler require that the client issue an additional explicit resource deallocation request once a job has terminated. Base-level clients would not know to do this and the result would be an incorrectly functioning system. Types of extensions I believe the following types of extensions are both necessary and sufficient to meet the needs of the HPC profile work: 5. * Addition of new WSDL operations. 6. * This is needed to support additional new functionality, such as the addition of suspend/resume operations. As long as base-level semantics aren't modified, this form of extension seems to be straight-forward. 7. * Addition of additional parameters to existing WSDL operations. 8. * As long as base-level semantics are maintained, this form of extension is also straight-forward. An example is adding a notification callback parameter to job creation requests. However, it is not clear whether all tooling can readily handle this form of "operation overloading". It may be better - from a pragmatic point-of-view - to define new WSDL operations (with appropriately defined names) that achieve the same effect. 9. * Support for array operations and other forms of batching. 10. * When 1000's of jobs are involved the efficiency gains of employing array operations for things like queries or abort requests are too significant to ignore. Hence a model in which every job must be interacted with on a strictly individual basis via an EPR is arguably unacceptable. 11. * One approach would be to simply add array operations alongside the corresponding individual operations, so that one can selectively interact with jobs (as well as things like data files) in either an "object-oriented" fashion or in "bulk-array" fashion. One could observe that the array operations enable the corresponding individual operations as a trivial special case, but this would arguably violate the principle of defining a minimalist base case and then employing only extensions (rather than replacements). 12. * Array operations are an example or a service-oriented rather than a resource-oriented form of interaction: clients send a single request to a job scheduler (service) that refers to an array of many resources, such as jobs. This raises the question of whether things like jobs should be referred to via EPRs or via unique "abstract" names that are independent of any given service's contact address. At a high level, the choice is unimportant since the client submitting an array operation request is simply using either one as a unique (and opaque) identifier for the relevant resource. On a pragmatic level one might argue that abstract names are easier and more efficient to deal with than EPRs since the receiving scheduler will need to parse EPRs to extract what is essentially the abstract name for each resource. (Using arrays of abstract names rather than arrays of EPRs is also more efficient from a size point-of-view.) 13. * If abstract names are used in array operations then it will necessary that individual operations return the abstract name and not just an EPR for a given resource, such as a job. If this approach is chosen then this implies that the base case design and implementation must return abstract names and not just EPRs for things like jobs. 14. * Extensions to state diagrams. 15. * Chris Smith is in the process of writing up this topic. 16. * Standardized extensions to things like resource definitions and other declarative definitions (e.g. about provisioning). 17. * The base use case assumes a small, fixed set of "standard" resources and other concepts (e.g. working directory) that may be described/requested. The simplest extension approach is to define additional specific "standard sets" that clients and services can refer to by their global name (e.g. the posix resource description set or the Windows resource description set) and of which they pick exactly one to use for any given interaction. 18. * The problem with this simplest form of extension is that it provides only a very crude form of extensibility with no notion of composition or incremental extension of existing definition sets. This is sufficient for very course-grained characterizations, such as "Windows environment" versus "Posix environment", but not for finer-grained resource extensions. An alternative is to define composable sets that cover specific "subjects" (e.g. GPUs). In the extreme, these sets could be of size 1. This implies that clients and services need to be able to deal with the power set of all possible meaningful combinations of these sets. As long as individual definitions are independent of each other (i.e. the semantics of specifying A is unchanged by specifying B in the same description) this isn't a big problem. Allowing the presence of different items in a description to affect each other's semantics is arguably a variation on modifying the base-level semantics of a design via some extension to the design and hence should be disallowed. 19. * If resource descriptions are used only for "matchmaking" against other resource descriptions then another approach is to allow arbitrary resource types whose semantics are not understood by the HPC infrastructure, which deals with them only as abstract entities whose names can be compared textually and whose associated values can be compared textually or numerically depending on their data type. It is important to understand that, whereas the "mechanical" aspects of an HPC infrastructure can mostly be built without having to know the semantics of these abstract resource types, their semantics must still be standardized and well-known at the level of the human beings using and programming the system. Both the descriptions of available computational resources and of client requests for reserving and using such resources must be specified in a manner that will cause the underlying HPC "matchmaking" infrastructure to do the right thing. This matchmaking approach is exemplified by systems such Condor's class ads system. 20. * It should be noted that a generalized matchmaking system is not a trivial thing to implement efficiently and hence one can reasonably imagine extensions based on any of the above approaches to extending resource (and other) definitions. 21. * Hierarchical and extended representations of information. 22. * XML infosets provide a very convenient way to represent extended descriptions of a particular piece of information. 23. * Another form of hierarchical information display shows up when multi-level scheduling systems are involved. In this case it may be desirable to represent information either in a form that hides the scheduling hierarchy or in a form that reflects it. Consider how to represent the list of compute nodes for a job running across multiple clusters: A flat view might list all compute nodes in an undifferentiated list. A hierarchical view might provide a list of clusters, each of which describes information about a cluster, including a list of the compute nodes in that cluster that the job is running on. Both views have their uses. XML infosets are convenient for encoding the syntax of either view, but an extension supporting information representation in these sorts of systems will also have to define the semantics of all allowed hierarchies. 24. * Decomposition of functionality into "micro" protocols. 25. * Micro protocols should reflect things that must occur at different times (e.g. resource reservation/allocation vs. resource use/job-execution) or that can be employed in a stand-alone manner (e.g. job execution vs. data transfer). The decomposition that seems relevant for the HPC use cases (i.e. are visible to clients) is the following: 26. * The base case involves interaction between a client and a scheduler for purposes of executing a job. 27. * A client may wish to independently reserve, or pre-allocate resources for later and/or guaranteed use. Note that this is different from simply submitting a job for execution to a scheduler that then queues the job for later execution - perhaps at a specific time requested by the client. For example, a meta-scheduler might wish to reserve resources so that it can make informed scheduling decisions about which "subsidiary" scheduler to send various jobs to. Similarly, a client might wish to reserve resources so as to run two separate jobs in succession to each other, with one job writing output to a scratch storage system and the second job reading that output as its input without having to worry that the data might have vanished during the interval that occurs between the execution of the two jobs. 28. * A client may wish to query a scheduler to learn what resources might be available to it, without actually laying claim to any resources as part of the query (let alone execute anything using those resources). Scheduling candidate set generators or matchmaking services such as Condor would want this functionality. 29. * A client may need to transfer specific data objects (e.g. files) to and from a system that is under the control of a job scheduling service. 30. * Micro protocols may have relationships to each other. For example, job execution will need to be able to accept a handle of some sort to resources that have already been allocated to the requesting client.