Inline. On Feb 27, 2015, at 6:40 AM, ruby.krishnasw...@orange.com<mailto:ruby.krishnasw...@orange.com> wrote:
My first suggestion: why don’t we set up call together with Ramki, Yathi, Debo, as soon as possible ? - How to go forward concretely with the 8 steps for the PoC (details within each step), o Including nova integration points - Identify “friction points” in the details above to resolve for beyond PoC Happy to have a call. But maybe it would be good to have something concrete to look at. Yali: could you outline how you envision the integration of Congress and solver-scheduler working in that same google doc? Maybe add a section on “Integration with Nova Solver-scheduler”? https://docs.google.com/document/d/1ksDilJYXV-5AXWON8PLMedDKr9NpS8VbT0jIy_MIEtI/edit# Tim: Where does the rest of the program originate? I’m not saying the entire LP program is generated from the Datalog constraints; some of it is generated by the solver independent of the Datalog. In the text, I gave the example of defining hMemUse[j]. Tim: The VM-placement engine does 2 things: (I) translates Datalog to LP and (ii) generates additional LP constraints. (Both work items could leverage any constraints that are builtin to a specific solver, e.g. the solver-scheduler. The point is that there are 2 distinct, conceptual origins of the LP constraints: those that represent the Datalog and those that codify the domain. Tim: Each domain-specific solver can do whatever it wants, so it’s not clear to me what the value of choosing a modeling language actually is—unless we want to build a library of common functionality that makes the construction of domain-specific engine (wrappers) easier. I’d prefer to spend our energy understanding whether the proposed workflow/interface works for a couple of different domain-specific policy engines OR to flush this one out and build it. ð The value of choosing a modeling language is related to how “best to automate translations” from Datalog constraints (to LP)? o We can have look for one unique way of generation, and not, “some of it is generated by the VM-placement engine solver independent of the Datalog”. o Datalog imposes most constraints (== policies) o Two constraints are not “policies” § A VM is allocated to only one host. § Host capacity is not exceeded. · Over subscription ð Otherwise what was your suggestion? As follows? o Use framework (extend) the nova-solver-scheduler currently implements (itself using PuLP). This framework specifies an API to write constraints and cost functions (in a domain specific way). Modifying this framework: § To read data in from DSE § To obtain the cost function from Datalog (e.g. minimize Y[host1]…) § To obtain Datalog constraints (e.g. <75% memory allocation constraint for hosts of special zone) o We need to specify the “format” for this? It will likely to be a string of the form (?) § “hMemUse[0] – 0.75*hMemCap[0] < 100*y[0], “ Memory allocation constraint on Host 0“, I wasn’t thinking of building another language for a front-end to the LP solver. I envisioned we’d use one that already exists like PuLP. I was just using notation in the doc that I thought everyone would understand (whether you know PuLP or not—I don’t). ð From your doc (page 5, section 4) warning(id) :- nova:host(id, name, service, zone, memory_capacity), legacy:special_zone(zone), ceilometer:statistics(id, "memory", avg, count, duration, durstart, durend, max, min, period, perstart, perend, sum, unit), avg > 0.75 * memory_capacity Notice that this is a soft constraint, identified by the use of warning instead of error. When compiling to LP, the VM-placement engine will attempt to minimize the number of rows in the warning table. That is, for each possible row r it will create a variable Y[r] and assign it True if the row is a warning and False otherwise The policy (warning) : when will it be evaluated? This should be done periodically? Then if the table has even one True entry, then the action should be to generate the LP, solve, activate the migrations etc. ð The “LP” cannot be generated when the VM-placement engine receives the policy snippet. It’s a good question about how often we re-generate/re-solve/re-migrate. I had assumed that would be built into the domain-specific engine, since it should know enough about the domain to make intelligent decisions. You could imagine doing that every time the engine received a new policy; every time it received new data; every 10 minutes; or every time new data arrives that is different enough. The engine could even re-generate/re-solve the LP and then decide whether the benefit of migrating to the new solution is worth the cost. And of course there’s a difference between the PoC and the long-term solution. I don’t know what the right answer is for the PoC. Suggestions? Tim Ruby De : Tim Hinrichs [mailto:thinri...@vmware.com] Envoyé : jeudi 26 février 2015 19:17 À : OpenStack Development Mailing List (not for usage questions) Objet : Re: [openstack-dev] [Congress][Delegation] Initial workflow design Inline. From: "ruby.krishnasw...@orange.com<mailto:ruby.krishnasw...@orange.com>" <ruby.krishnasw...@orange.com<mailto:ruby.krishnasw...@orange.com>> Reply-To: "OpenStack Development Mailing List (not for usage questions)" <openstack-dev@lists.openstack.org<mailto:openstack-dev@lists.openstack.org>> Date: Wednesday, February 25, 2015 at 8:53 AM To: "OpenStack Development Mailing List (not for usage questions)" <openstack-dev@lists.openstack.org<mailto:openstack-dev@lists.openstack.org>> Subject: Re: [openstack-dev] [Congress][Delegation] Initial workflow design Hi Tim, All, 1) Step 3: The VM-placement engine is also a “datalog engine” . Right? When policies are delegated: when policies are inserted? When the VM-placement engine has already registered itself all policies are given to it? “In our example, this would mean the domain-specific policy engine executes the following API call over the DSE” ð “domain-agnostic” …. Done. 2) Step 4: Ok But finally: if Congress will likely “delegate” Not sure what you’re suggesting here. 3) Step 5: Compilation of subpolicy to LP in VM-placement engine For the PoC, it is likely that the LP program ( in PuLP or some other ML) is *not* completely generated by compiler/translator. ð Right? Where does the rest of the program originate? I’m not saying the entire LP program is generated from the Datalog constraints; some of it is generated by the solver independent of the Datalog. In the text, I gave the example of defining hMemUse[j]. You also indicate that some category of constraints (“the LP solver doesn’t know what the relationship between assign[i][j], hMemUse[j], and vMemUse[i] actually is, so the VM-placement engine must also include constraints”) . These constraints must be “explicitly” written? (e.g. max_ram_allocation etc that are constraints used in the solver-scheduler’s package). The VM-placement engine does 2 things: (I) translates Datalog to LP and (ii) generates additional LP constraints. (Both work items could leverage any constraints that are builtin to a specific solver, e.g. the solver-scheduler. The point is that there are 2 distinct, conceptual origins of the LP constraints: those that represent the Datalog and those that codify the domain. So what “parts” will be generated: Cost function : Constraint from Policy : memory usage < 75% Then the rest should be “filled” up? Could we convene on an intermediary “modeling language”? @Yathi: do you think we could use some thing like AMPL ? Is this proprietary? A detail: the example “Y[host1] = hMemUse[host1] > 0.75 * hMemCap[host1]” ð To be changed to a linear form (mi – Mi > 0 then Yi = 1 else Yi = 0) so something like (mi – Mi) < 100 yi Each domain-specific solver can do whatever it wants, so it’s not clear to me what the value of choosing a modeling language actually is—unless we want to build a library of common functionality that makes the construction of domain-specific engine (wrappers) easier. I’d prefer to spend our energy understanding whether the proposed workflow/interface works for a couple of different domain-specific policy engines OR to flush this one out and build it. 4) Step 6: This is completely internal to the VM-placement engine (and we could say this is “transparent” to Congress) We should allow configuration of a solver (this itself could be a policy ☺ ) How to invoke the solver API ? The domain-specific placement engine could send out to DSE (action_handler: data)? I had always envisioned the solver being just a library of code—not an entity that sits on the DSE itself. 3) Step 7 : Perform the migrations (according to the assignments computed in the step 6) This part invokes OpenStack API (to perform migrations). We may suppose that there are services implementing “action handlers”? It can listen on the DSE and execute the action. That interface is supposed to exist by the Kilo release. I’ll check up on the progress. 5) Nova tables to use Policy warning(id) :- nova:host(id, name, service, zone, memory_capacity), legacy:special_zone(zone), ceilometer:statistics(id, "memory", avg, count, duration, durstart, durend, max, min, period, perstart, perend, sum, unit), avg > 0.75 * memory_capacity I believe that ceilometer gives usage of VMs and not hosts. The host table (ComputeNode table) should give current used capacity. Good to know. 6) One of the issues highlighted in OpenStack (scheduler) and also elsewhere (e.g. Omega scheduler by google) is : Reading “host utilization” state from the data bases and DB (nova:host table) updates and overhead of maintaining in-memory state uptodate. ð This is expensive and current nova-scheduler does face this issue (many blogs/discussions). While the first goal is a PoC, this will likely become a concern in terms of adoption. So you’re saying we won’t have fresh enough data to make policy decisions? If the data changes so frequently that we can’t get an accurate view, then I’m guessing we shouldn’t be migrating based on that data anyway. Could you point me to some of these discussions? 7) While in this document you have changed the “example” policy, could we drill down the set of policies for the PoC (the server under utilization ?) ð As a reference Sure. The only reason I chose this policy was because it doesn’t have aggregation. I’m guessing we’ll want to avoid aggregation for the POC because we don’t yet have it in Congress, and it complicates the problem of translating Datalog to LP substantially. Tim Ruby De : Yathiraj Udupi (yudupi) [mailto:yud...@cisco.com] Envoyé : mardi 24 février 2015 20:01 À : OpenStack Development Mailing List (not for usage questions); Tim Hinrichs Cc : Debo Dutta (dedutta) Objet : Re: [openstack-dev] [Congress][Delegation] Initial workflow design Hi Tim, Thanks for your updated doc on Delegation from Congress to a domain-specific policy engine, in this case, you are planning to build a LP-based VM-Placement engine to be the domain specific policy engine. I agree your main goal is to first get the delegation interface sorted out. It will be good so that external services (like Solver-Scheduler) can also easily integrate to the delegation model. From the Solver-Scheduler point of view, we would actually want to start working on a PoC effort to start integrating Congress and the Solver-Scheduler. We believe rather than pushing this effort to a long-term, it would add value to both the Solver Scheduler effort, as well as the Congress effort to try some early integration now, as most of the LP solver work for VM placements is ready available now in Solver scheduler, and we need to spend some time thinking about translating your domain-agnostic policy to constraints that the Solver scheduler can use. I would definitely need your help from the Congress interfaces and I hope you will share your early interfaces for the delegation, so I can start the effort from the Solver scheduler side for integration. I will reach out to you to get some initial help for integration w.r.t. Congress, and also keep you posted about the progress from our side. Thanks, Yathi. On 2/23/15, 11:28 AM, "Tim Hinrichs" <thinri...@vmware.com<mailto:thinri...@vmware.com>> wrote: Hi all, I made a heavy editing pass of the Delegation google doc, incorporating many of your comments and my latest investigations into VM-placement. I left the old stuff in place at the end of the doc and put the new stuff at the top. My goal was to propose an end-to-end workflow for a PoC that we could put together quickly to help us explore the delegation interface. We should iterate on this design until we have something that we think is workable. And by all means pipe up if you think we need a totally different starting point to begin the iteration. (BTW I'm thinking of the integration with solver-scheduler as a long-term solution to VM-placement, once we get the delegation interface sorted out.) https://docs.google.com/document/d/1ksDilJYXV-5AXWON8PLMedDKr9NpS8VbT0jIy_MIEtI/edit#<https://urldefense.proofpoint.com/v2/url?u=https-3A__docs.google.com_document_d_1ksDilJYXV-2D5AXWON8PLMedDKr9NpS8VbT0jIy-5FMIEtI_edit&d=AwMFAw&c=Sqcl0Ez6M0X8aeM67LKIiDJAXVeAw-YihVMNtXt-uEs&r=B6BWd4kFfgOzAREgThxkmTZKy7dDXE2-eBAmL0PBK7s&m=kF8jMOpogOhk8MJWvNMKJC3PiNImxWpZeD2o642YM2s&s=8PV5EW-kz8Q9aP9riFbIjJXJNZXchx2NsL-Z3Y7E5Vg&e=> Tim _________________________________________________________________________________________________________________________ Ce message et ses pieces jointes peuvent contenir des informations confidentielles ou privilegiees et ne doivent donc pas etre diffuses, exploites ou copies sans autorisation. Si vous avez recu ce message par erreur, veuillez le signaler a l'expediteur et le detruire ainsi que les pieces jointes. Les messages electroniques etant susceptibles d'alteration, Orange decline toute responsabilite si ce message a ete altere, deforme ou falsifie. Merci. This message and its attachments may contain confidential or privileged information that may be protected by law; they should not be distributed, used or copied without authorisation. If you have received this email in error, please notify the sender and delete this message and its attachments. As emails may be altered, Orange is not liable for messages that have been modified, changed or falsified. Thank you. _________________________________________________________________________________________________________________________ Ce message et ses pieces jointes peuvent contenir des informations confidentielles ou privilegiees et ne doivent donc pas etre diffuses, exploites ou copies sans autorisation. Si vous avez recu ce message par erreur, veuillez le signaler a l'expediteur et le detruire ainsi que les pieces jointes. Les messages electroniques etant susceptibles d'alteration, Orange decline toute responsabilite si ce message a ete altere, deforme ou falsifie. Merci. This message and its attachments may contain confidential or privileged information that may be protected by law; they should not be distributed, used or copied without authorisation. If you have received this email in error, please notify the sender and delete this message and its attachments. As emails may be altered, Orange is not liable for messages that have been modified, changed or falsified. Thank you. __________________________________________________________________________ OpenStack Development Mailing List (not for usage questions) Unsubscribe: openstack-dev-requ...@lists.openstack.org<mailto:openstack-dev-requ...@lists.openstack.org>?subject:unsubscribe http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
__________________________________________________________________________ OpenStack Development Mailing List (not for usage questions) Unsubscribe: openstack-dev-requ...@lists.openstack.org?subject:unsubscribe http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
