Re: [openstack-dev] [Congress][Delegation] Google doc for working notes
Hi Ruby, I was envisioning the VM-placement engine choosing the hard-coded implementation or converting to LP. There are 2 places that logic could go: in the VM-placement engine wrapper that runs on the DSE message bus, or within the VM-placement engine itself (assuming the two are different). I think we’re still trying to figure out which one (and the initial PoC and the long-term solution may be different). Right now I’m thinking that the VM-placement engine wrapper running on the DSE bus should subscribe to whatever data it needs. Tim On Feb 16, 2015, at 9:05 AM, ruby.krishnasw...@orange.commailto:ruby.krishnasw...@orange.com ruby.krishnasw...@orange.commailto:ruby.krishnasw...@orange.com wrote: Hi Tim What I’d like to see is the best of both worlds. Users write Datalog policies describing whatever VM-placement policy they want. If the policy they’ve written is on the solver-scheduler’s list of options, we use the hard-coded implementation, but if the policy isn’t on that list we translate directly to LP. ð How (calling the hard-coded implementation) ? o Through the message bus? ð Is it Congress that will send out the data or should each implementation (of a policy) read it in directly? Ruby De : Tim Hinrichs [mailto:thinri...@vmware.com] Envoyé : jeudi 12 février 2015 19:03 À : OpenStack Development Mailing List (not for usage questions) Objet : Re: [openstack-dev] [Congress][Delegation] Google doc for working notes Hi Debo and Yathiraj, I took a third look at the solver-scheduler docs and code with your comments in mind. A few things jumped out. 1) Choice of LP solver. I see solver-scheduler uses Pulp, which was on the Congress short list as well. So we’re highly aligned on the choice of underlying solver. 2) User control over VM-placement. To choose the criteria for VM-placement, the solver-scheduler user picks from a list of predefined options, e.g. ActiveHostConstraint, MaxRamAllocationPerHostConstraint. We’re investigating a slightly different approach, where the user defines the criteria for VM-placement by writing any policy they like in Datalog. Under the hood we then convert that Datalog to an LP problem. From the developer’s perspective, with the Congress approach we don’t attempt to anticipate the different policies the user might want and write code for each policy; instead, we as developers write a translator from Datalog to LP. From the user’s perspective, the difference is that if the option they want isn’t on the solver-scheduler's list, they’re out of luck or need to write the code themselves. But with the Congress approach, they can write any VM-placement policy they like. What I’d like to see is the best of both worlds. Users write Datalog policies describing whatever VM-placement policy they want. If the policy they’ve written is on the solver-scheduler’s list of options, we use the hard-coded implementation, but if the policy isn’t on that list we translate directly to LP. This approach gives us the ability to write custom code to handle common cases while at the same time letting users write whatever policy they like. 3) API and architecture. Today the solver-scheduler's VM-placement policy is defined at config-time (i.e. not run-time). Am I correct that this limitation is only because there’s no API call to set the solver-scheduler’s policy? Or is there some other reason the policy is set at config-time? Congress policies change at runtime, so we’ll definitely need a VM-placement engine whose policy can be changed at run-time as well. If we focus on just migration (and not provisioning), we can build a VM-placement engine that sits outside of Nova that has an API call that allows us to set policy at runtime. We can also set up that engine to get data updates that influence the policy. We were planning on creating this kind of VM-placement engine within Congress as a node on the DSE (our message bus). This is convenient because all nodes on the DSE run in their own thread, any node on the DSE can subscribe to any data from any other node (e.g. ceilometer’s data), and the algorithms for translating Datalog to LP look to be quite similar to the algorithms we’re using in our domain-agnostic policy engine. Tim On Feb 11, 2015, at 4:50 PM, Debojyoti Dutta ddu...@gmail.commailto:ddu...@gmail.com wrote: Hi Tim: moving our thread to the mailer. Excited to collaborate! From: Debo~ Dutta dedu...@cisco.commailto:dedu...@cisco.com Date: Wednesday, February 11, 2015 at 4:48 PM To: Tim Hinrichs thinri...@vmware.commailto:thinri...@vmware.com Cc: Yathiraj Udupi (yudupi) yud...@cisco.commailto:yud...@cisco.com, Gokul B Kandiraju go...@us.ibm.commailto:go...@us.ibm.com, Prabhakar Kudva ku...@us.ibm.commailto:ku...@us.ibm.com, ruby.krishnasw...@orange.commailto:ruby.krishnasw...@orange.com ruby.krishnasw...@orange.commailto:ruby.krishnasw...@orange.com, dilik...@in.ibm.commailto:dilik
Re: [openstack-dev] [Congress][Delegation] Google doc for working notes
Hi Tim What I'd like to see is the best of both worlds. Users write Datalog policies describing whatever VM-placement policy they want. If the policy they've written is on the solver-scheduler's list of options, we use the hard-coded implementation, but if the policy isn't on that list we translate directly to LP. ð How (calling the hard-coded implementation) ? o Through the message bus? ð Is it Congress that will send out the data or should each implementation (of a policy) read it in directly? Ruby De : Tim Hinrichs [mailto:thinri...@vmware.com] Envoyé : jeudi 12 février 2015 19:03 À : OpenStack Development Mailing List (not for usage questions) Objet : Re: [openstack-dev] [Congress][Delegation] Google doc for working notes Hi Debo and Yathiraj, I took a third look at the solver-scheduler docs and code with your comments in mind. A few things jumped out. 1) Choice of LP solver. I see solver-scheduler uses Pulp, which was on the Congress short list as well. So we're highly aligned on the choice of underlying solver. 2) User control over VM-placement. To choose the criteria for VM-placement, the solver-scheduler user picks from a list of predefined options, e.g. ActiveHostConstraint, MaxRamAllocationPerHostConstraint. We're investigating a slightly different approach, where the user defines the criteria for VM-placement by writing any policy they like in Datalog. Under the hood we then convert that Datalog to an LP problem. From the developer's perspective, with the Congress approach we don't attempt to anticipate the different policies the user might want and write code for each policy; instead, we as developers write a translator from Datalog to LP. From the user's perspective, the difference is that if the option they want isn't on the solver-scheduler's list, they're out of luck or need to write the code themselves. But with the Congress approach, they can write any VM-placement policy they like. What I'd like to see is the best of both worlds. Users write Datalog policies describing whatever VM-placement policy they want. If the policy they've written is on the solver-scheduler's list of options, we use the hard-coded implementation, but if the policy isn't on that list we translate directly to LP. This approach gives us the ability to write custom code to handle common cases while at the same time letting users write whatever policy they like. 3) API and architecture. Today the solver-scheduler's VM-placement policy is defined at config-time (i.e. not run-time). Am I correct that this limitation is only because there's no API call to set the solver-scheduler's policy? Or is there some other reason the policy is set at config-time? Congress policies change at runtime, so we'll definitely need a VM-placement engine whose policy can be changed at run-time as well. If we focus on just migration (and not provisioning), we can build a VM-placement engine that sits outside of Nova that has an API call that allows us to set policy at runtime. We can also set up that engine to get data updates that influence the policy. We were planning on creating this kind of VM-placement engine within Congress as a node on the DSE (our message bus). This is convenient because all nodes on the DSE run in their own thread, any node on the DSE can subscribe to any data from any other node (e.g. ceilometer's data), and the algorithms for translating Datalog to LP look to be quite similar to the algorithms we're using in our domain-agnostic policy engine. Tim On Feb 11, 2015, at 4:50 PM, Debojyoti Dutta ddu...@gmail.commailto:ddu...@gmail.com wrote: Hi Tim: moving our thread to the mailer. Excited to collaborate! From: Debo~ Dutta dedu...@cisco.commailto:dedu...@cisco.com Date: Wednesday, February 11, 2015 at 4:48 PM To: Tim Hinrichs thinri...@vmware.commailto:thinri...@vmware.com Cc: Yathiraj Udupi (yudupi) yud...@cisco.commailto:yud...@cisco.com, Gokul B Kandiraju go...@us.ibm.commailto:go...@us.ibm.com, Prabhakar Kudva ku...@us.ibm.commailto:ku...@us.ibm.com, ruby.krishnasw...@orange.commailto:ruby.krishnasw...@orange.com ruby.krishnasw...@orange.commailto:ruby.krishnasw...@orange.com, dilik...@in.ibm.commailto:dilik...@in.ibm.com dilik...@in.ibm.commailto:dilik...@in.ibm.com, Norival Figueira nfigu...@brocade.commailto:nfigu...@brocade.com, Ramki Krishnan r...@brocade.commailto:r...@brocade.com, Xinyuan Huang (xinyuahu) xinyu...@cisco.commailto:xinyu...@cisco.com, Rishabh Jain -X (rishabja - AAP3 INC at Cisco) risha...@cisco.commailto:risha...@cisco.com Subject: Re: Nova solver scheduler and Congress Hi Tim To address your particular questions: 1. translate some policy language into constraints for the LP/CVP and we had left that to congress hoping to integrate when the policy efforts in openstack were ready (our initial effort was pre congress) 2. For migrations: we are currently doing that - its about incremental
Re: [openstack-dev] [Congress][Delegation] Google doc for working notes
Tim Wanted to clarify a bit. As I have mentioned before: Solver scheduler is work done before this work (Datalog-constraints) but we had kept it very generic to be integrated with something like congress. In fact Ramki (who was one of the members of the original thread when you reached out to us) joined us to in talk in Atlanta where we described some of the same use cases using PULP congress was still ramping up then. We were not aware of the Datalog-constraints work that you guys were doing, else we would have joined hands before. The question is this: going forward, how do build this cool stuff together in the community? I am hoping the scheduler folks will be very excited too! debo On Thu, Feb 12, 2015 at 11:27 AM, Yathiraj Udupi (yudupi) yud...@cisco.com wrote: Hi Tim, Thanks for your response. Excited too to extend the collaboration and ensure there is no need to duplicate effort in the open source community. My responses inline. 1) Choice of LP solver. I see solver-scheduler uses Pulp, which was on the Congress short list as well. So we’re highly aligned on the choice of underlying solver. YATHI - This makes me wonder why can’t we easily adapt the solver-scheduler to your needs, rather than duplicating the effort! 2) User control over VM-placement. To choose the criteria for VM-placement, the solver-scheduler user picks from a list of predefined options, e.g. ActiveHostConstraint, MaxRamAllocationPerHostConstraint. We’re investigating a slightly different approach, where the user defines the criteria for VM-placement by writing any policy they like in Datalog. Under the hood we then convert that Datalog to an LP problem. From the developer’s perspective, with the Congress approach we don’t attempt to anticipate the different policies the user might want and write code for each policy; instead, we as developers write a translator from Datalog to LP. From the user’s perspective, the difference is that if the option they want isn’t on the solver-scheduler's list, they’re out of luck or need to write the code themselves. But with the Congress approach, they can write any VM-placement policy they like. What I’d like to see is the best of both worlds. Users write Datalog policies describing whatever VM-placement policy they want. If the policy they’ve written is on the solver-scheduler’s list of options, we use the hard-coded implementation, but if the policy isn’t on that list we translate directly to LP. This approach gives us the ability to write custom code to handle common cases while at the same time letting users write whatever policy they like. YATHI - The idea of providing some default constraint classes in Solver Scheduler was to enable easy pluggability for various placement policy scenarios. We can easily add a custom constraint class in solver scheduler, that enables adding additional constraints at runtime (PulP model or any other models we can use and support). It will just take in any external policy (say Datalog in Congress example), and it can easily add those set of resulting translated constraints via this custom constraint builder class. This is something we can definitely add value to solver scheduler by implementing and adding here. 3) API and architecture. Today the solver-scheduler's VM-placement policy is defined at config-time (i.e. not run-time). Am I correct that this limitation is only because there’s no API call to set the solver-scheduler’s policy? Or is there some other reason the policy is set at config-time? Congress policies change at runtime, so we’ll definitely need a VM-placement engine whose policy can be changed at run-time as well. YATHI - We have working code to set VM placement policies at run-time to dynamically select the constraint or cost classes to use. It is yet to upstreamed to solver scheduler stackforge repo, but will be soon. But yeah I agree with you, this is definitely needed for any policy-driven VM placement engine, as the policies are dynamic. Short answer, yes solver scheduler has abilities to support this. If we focus on just migration (and not provisioning), we can build a VM-placement engine that sits outside of Nova that has an API call that allows us to set policy at runtime. We can also set up that engine to get data updates that influence the policy. We were planning on creating this kind of VM-placement engine within Congress as a node on the DSE (our message bus). This is convenient because all nodes on the DSE run in their own thread, any node on the DSE can subscribe to any data from any other node (e.g. ceilometer’s data), and the algorithms for translating Datalog to LP look to be quite similar to the algorithms we’re using in our domain-agnostic policy engine. YATHI – The entire scheduling community in Nova is planning on an external scheduler (Gantt), and we are pitching solver scheduler also
Re: [openstack-dev] [Congress][Delegation] Google doc for working notes
Hello Debo/Tim My understanding is that with Congress things like filters (e.g. anti-affinity or other aggregates) will be replaced to be written as policies with Datalog. Goals (a Policy), Constraints (policies in Congress) will also get translated to (for example) linear programs in some modelling language (e.g. PuLP). So this is likely to be a major change to the scheduler? Ruby De : Debojyoti Dutta [mailto:ddu...@gmail.com] Envoyé : vendredi 13 février 2015 14:06 À : OpenStack Development Mailing List (not for usage questions) Objet : Re: [openstack-dev] [Congress][Delegation] Google doc for working notes Tim Wanted to clarify a bit. As I have mentioned before: Solver scheduler is work done before this work (Datalog-constraints) but we had kept it very generic to be integrated with something like congress. In fact Ramki (who was one of the members of the original thread when you reached out to us) joined us to in talk in Atlanta where we described some of the same use cases using PULP congress was still ramping up then. We were not aware of the Datalog-constraints work that you guys were doing, else we would have joined hands before. The question is this: going forward, how do build this cool stuff together in the community? I am hoping the scheduler folks will be very excited too! debo On Thu, Feb 12, 2015 at 11:27 AM, Yathiraj Udupi (yudupi) yud...@cisco.commailto:yud...@cisco.com wrote: Hi Tim, Thanks for your response. Excited too to extend the collaboration and ensure there is no need to duplicate effort in the open source community. My responses inline. 1) Choice of LP solver. I see solver-scheduler uses Pulp, which was on the Congress short list as well. So we’re highly aligned on the choice of underlying solver. YATHI - This makes me wonder why can’t we easily adapt the solver-scheduler to your needs, rather than duplicating the effort! 2) User control over VM-placement. To choose the criteria for VM-placement, the solver-scheduler user picks from a list of predefined options, e.g. ActiveHostConstraint, MaxRamAllocationPerHostConstraint. We’re investigating a slightly different approach, where the user defines the criteria for VM-placement by writing any policy they like in Datalog. Under the hood we then convert that Datalog to an LP problem. From the developer’s perspective, with the Congress approach we don’t attempt to anticipate the different policies the user might want and write code for each policy; instead, we as developers write a translator from Datalog to LP. From the user’s perspective, the difference is that if the option they want isn’t on the solver-scheduler's list, they’re out of luck or need to write the code themselves. But with the Congress approach, they can write any VM-placement policy they like. What I’d like to see is the best of both worlds. Users write Datalog policies describing whatever VM-placement policy they want. If the policy they’ve written is on the solver-scheduler’s list of options, we use the hard-coded implementation, but if the policy isn’t on that list we translate directly to LP. This approach gives us the ability to write custom code to handle common cases while at the same time letting users write whatever policy they like. YATHI - The idea of providing some default constraint classes in Solver Scheduler was to enable easy pluggability for various placement policy scenarios. We can easily add a custom constraint class in solver scheduler, that enables adding additional constraints at runtime (PulP model or any other models we can use and support). It will just take in any external policy (say Datalog in Congress example), and it can easily add those set of resulting translated constraints via this custom constraint builder class. This is something we can definitely add value to solver scheduler by implementing and adding here. 3) API and architecture. Today the solver-scheduler's VM-placement policy is defined at config-time (i.e. not run-time). Am I correct that this limitation is only because there’s no API call to set the solver-scheduler’s policy? Or is there some other reason the policy is set at config-time? Congress policies change at runtime, so we’ll definitely need a VM-placement engine whose policy can be changed at run-time as well. YATHI - We have working code to set VM placement policies at run-time to dynamically select the constraint or cost classes to use. It is yet to upstreamed to solver scheduler stackforge repo, but will be soon. But yeah I agree with you, this is definitely needed for any policy-driven VM placement engine, as the policies are dynamic. Short answer, yes solver scheduler has abilities to support this. If we focus on just migration (and not provisioning), we can build a VM-placement engine that sits outside of Nova that has an API call that allows us to set policy at runtime. We can also set up that engine
Re: [openstack-dev] [Congress][Delegation] Google doc for working notes
Hi Debo and Yathi, We’re completely on the same page here. We’ve known about the solver-scheduler for a while now (I even attended your Atlanta talk), and I always expected Congress would integrate with it. As you say, now it’s a matter of getting down to the details. A bit on the context… The current problem we’re working on in Congress is how we might delegate responsibility for policy enforcement to domain-specific policy engines, and a number of people were interested in integrating with a VM-placement engine. We started looking at the solver-scheduler (the obvious first choice), hence this dialog. The notes in the google doc are just me trying to understand the problem of delegation to a VM-placement engine by working through the problem end-to-end. (I’ve not worked with LP or VM-placement much, so my notes are there to help me grapple a bit with the domain for the first time.) How we build a PoC is something we haven’t started to discuss. So you’re joining the discussion at the right time. The more of that PoC we can build by leveraging solver-scheduler, the better. More detailed comments inline. On Feb 13, 2015, at 5:05 AM, Debojyoti Dutta ddu...@gmail.commailto:ddu...@gmail.com wrote: Tim Wanted to clarify a bit. As I have mentioned before: Solver scheduler is work done before this work (Datalog-constraints) but we had kept it very generic to be integrated with something like congress. In fact Ramki (who was one of the members of the original thread when you reached out to us) joined us to in talk in Atlanta where we described some of the same use cases using PULP congress was still ramping up then. We were not aware of the Datalog-constraints work that you guys were doing, else we would have joined hands before. The question is this: going forward, how do build this cool stuff together in the community? I am hoping the scheduler folks will be very excited too! debo On Thu, Feb 12, 2015 at 11:27 AM, Yathiraj Udupi (yudupi) yud...@cisco.commailto:yud...@cisco.com wrote: Hi Tim, Thanks for your response. Excited too to extend the collaboration and ensure there is no need to duplicate effort in the open source community. My responses inline. 1) Choice of LP solver. I see solver-scheduler uses Pulp, which was on the Congress short list as well. So we’re highly aligned on the choice of underlying solver. YATHI - This makes me wonder why can’t we easily adapt the solver-scheduler to your needs, rather than duplicating the effort! My primary goal is to build an architecture that makes it easy to integrate with domain-specific policy engines (like compute or networking). What I’m also hearing is that people are interested in building *new* domain-specific policy engines within the Congress framework and/or expanding the functionality of the Congress policy engine itself to include optimization technology. In both cases, we would need a library for solving optimization problems. Oliver (CC’ed) has proposed adding such a library to Congress. Solver-scheduler already has such a library, so it would be great if we could all brainstorm about how to make optimization technology easy to use for people writing domain-specific policy engines, without reinventing the wheel. https://blueprints.launchpad.net/congress/+spec/rule-x 2) User control over VM-placement. To choose the criteria for VM-placement, the solver-scheduler user picks from a list of predefined options, e.g. ActiveHostConstraint, MaxRamAllocationPerHostConstraint. We’re investigating a slightly different approach, where the user defines the criteria for VM-placement by writing any policy they like in Datalog. Under the hood we then convert that Datalog to an LP problem. From the developer’s perspective, with the Congress approach we don’t attempt to anticipate the different policies the user might want and write code for each policy; instead, we as developers write a translator from Datalog to LP. From the user’s perspective, the difference is that if the option they want isn’t on the solver-scheduler's list, they’re out of luck or need to write the code themselves. But with the Congress approach, they can write any VM-placement policy they like. What I’d like to see is the best of both worlds. Users write Datalog policies describing whatever VM-placement policy they want. If the policy they’ve written is on the solver-scheduler’s list of options, we use the hard-coded implementation, but if the policy isn’t on that list we translate directly to LP. This approach gives us the ability to write custom code to handle common cases while at the same time letting users write whatever policy they like. YATHI - The idea of providing some default constraint classes in Solver Scheduler was to enable easy pluggability for various placement policy scenarios. We can easily add a custom constraint class in solver scheduler, that enables adding
Re: [openstack-dev] [Congress][Delegation] Google doc for working notes
Hi Tim, Glad to collaborate and work towards nailing down the details. Yeah in terms of policy enforcement from Congress, it makes sense to delegate to domain-specific policy engines. It will be good to go through this PoC and to start thinking about the integration points of Congress with Solver scheduler, with a good set of APIs supported from both sides. Some more comments inline to your questions: On 2/13/15, 9:35 AM, Tim Hinrichs thinri...@vmware.commailto:thinri...@vmware.com wrote: Hi Debo and Yathi, We’re completely on the same page here. We’ve known about the solver-scheduler for a while now (I even attended your Atlanta talk), and I always expected Congress would integrate with it. As you say, now it’s a matter of getting down to the details. A bit on the context… The current problem we’re working on in Congress is how we might delegate responsibility for policy enforcement to domain-specific policy engines, and a number of people were interested in integrating with a VM-placement engine. We started looking at the solver-scheduler (the obvious first choice), hence this dialog. The notes in the google doc are just me trying to understand the problem of delegation to a VM-placement engine by working through the problem end-to-end. (I’ve not worked with LP or VM-placement much, so my notes are there to help me grapple a bit with the domain for the first time.) How we build a PoC is something we haven’t started to discuss. So you’re joining the discussion at the right time. The more of that PoC we can build by leveraging solver-scheduler, the better. More detailed comments inline. On Feb 13, 2015, at 5:05 AM, Debojyoti Dutta ddu...@gmail.commailto:ddu...@gmail.com wrote: Tim Wanted to clarify a bit. As I have mentioned before: Solver scheduler is work done before this work (Datalog-constraints) but we had kept it very generic to be integrated with something like congress. In fact Ramki (who was one of the members of the original thread when you reached out to us) joined us to in talk in Atlanta where we described some of the same use cases using PULP congress was still ramping up then. We were not aware of the Datalog-constraints work that you guys were doing, else we would have joined hands before. The question is this: going forward, how do build this cool stuff together in the community? I am hoping the scheduler folks will be very excited too! debo On Thu, Feb 12, 2015 at 11:27 AM, Yathiraj Udupi (yudupi) yud...@cisco.commailto:yud...@cisco.com wrote: Hi Tim, Thanks for your response. Excited too to extend the collaboration and ensure there is no need to duplicate effort in the open source community. My responses inline. 1) Choice of LP solver. I see solver-scheduler uses Pulp, which was on the Congress short list as well. So we’re highly aligned on the choice of underlying solver. YATHI - This makes me wonder why can’t we easily adapt the solver-scheduler to your needs, rather than duplicating the effort! My primary goal is to build an architecture that makes it easy to integrate with domain-specific policy engines (like compute or networking). What I’m also hearing is that people are interested in building *new* domain-specific policy engines within the Congress framework and/or expanding the functionality of the Congress policy engine itself to include optimization technology. In both cases, we would need a library for solving optimization problems. Oliver (CC’ed) has proposed adding such a library to Congress. Solver-scheduler already has such a library, so it would be great if we could all brainstorm about how to make optimization technology easy to use for people writing domain-specific policy engines, without reinventing the wheel. https://blueprints.launchpad.net/congress/+spec/rule-x YATHI: It is an interesting thought in this direction. However it is good to build an architecture with easy integration of Congress to separate domain-specific engines in compute, networking, and storage.I feel some of these policy validation/enforcement workflows are sometimes best within the domain-specific engines like VM placement/scheduling for example. But definitely optimization technology is a good choice for this kind of problems. 2) User control over VM-placement. To choose the criteria for VM-placement, the solver-scheduler user picks from a list of predefined options, e.g. ActiveHostConstraint, MaxRamAllocationPerHostConstraint. We’re investigating a slightly different approach, where the user defines the criteria for VM-placement by writing any policy they like in Datalog. Under the hood we then convert that Datalog to an LP problem. From the developer’s perspective, with the Congress approach we don’t attempt to anticipate the different policies the user might want and write code for each policy; instead, we as developers write a translator from Datalog to LP. From the user’s
Re: [openstack-dev] [Congress][Delegation] Google doc for working notes
Hi Tim, Thanks for your response. Excited too to extend the collaboration and ensure there is no need to duplicate effort in the open source community. My responses inline. 1) Choice of LP solver. I see solver-scheduler uses Pulp, which was on the Congress short list as well. So we’re highly aligned on the choice of underlying solver. YATHI - This makes me wonder why can’t we easily adapt the solver-scheduler to your needs, rather than duplicating the effort! 2) User control over VM-placement. To choose the criteria for VM-placement, the solver-scheduler user picks from a list of predefined options, e.g. ActiveHostConstraint, MaxRamAllocationPerHostConstraint. We’re investigating a slightly different approach, where the user defines the criteria for VM-placement by writing any policy they like in Datalog. Under the hood we then convert that Datalog to an LP problem. From the developer’s perspective, with the Congress approach we don’t attempt to anticipate the different policies the user might want and write code for each policy; instead, we as developers write a translator from Datalog to LP. From the user’s perspective, the difference is that if the option they want isn’t on the solver-scheduler's list, they’re out of luck or need to write the code themselves. But with the Congress approach, they can write any VM-placement policy they like. What I’d like to see is the best of both worlds. Users write Datalog policies describing whatever VM-placement policy they want. If the policy they’ve written is on the solver-scheduler’s list of options, we use the hard-coded implementation, but if the policy isn’t on that list we translate directly to LP. This approach gives us the ability to write custom code to handle common cases while at the same time letting users write whatever policy they like. YATHI - The idea of providing some default constraint classes in Solver Scheduler was to enable easy pluggability for various placement policy scenarios. We can easily add a custom constraint class in solver scheduler, that enables adding additional constraints at runtime (PulP model or any other models we can use and support). It will just take in any external policy (say Datalog in Congress example), and it can easily add those set of resulting translated constraints via this custom constraint builder class. This is something we can definitely add value to solver scheduler by implementing and adding here. 3) API and architecture. Today the solver-scheduler's VM-placement policy is defined at config-time (i.e. not run-time). Am I correct that this limitation is only because there’s no API call to set the solver-scheduler’s policy? Or is there some other reason the policy is set at config-time? Congress policies change at runtime, so we’ll definitely need a VM-placement engine whose policy can be changed at run-time as well. YATHI - We have working code to set VM placement policies at run-time to dynamically select the constraint or cost classes to use. It is yet to upstreamed to solver scheduler stackforge repo, but will be soon. But yeah I agree with you, this is definitely needed for any policy-driven VM placement engine, as the policies are dynamic. Short answer, yes solver scheduler has abilities to support this. If we focus on just migration (and not provisioning), we can build a VM-placement engine that sits outside of Nova that has an API call that allows us to set policy at runtime. We can also set up that engine to get data updates that influence the policy. We were planning on creating this kind of VM-placement engine within Congress as a node on the DSE (our message bus). This is convenient because all nodes on the DSE run in their own thread, any node on the DSE can subscribe to any data from any other node (e.g. ceilometer’s data), and the algorithms for translating Datalog to LP look to be quite similar to the algorithms we’re using in our domain-agnostic policy engine. YATHI – The entire scheduling community in Nova is planning on an external scheduler (Gantt), and we are pitching solver scheduler also as a stand-alone placement engine a scheduler as a service. Nova integration is just to ensure it fits within the Nova workflow. I am not quite familiar with the DSE architecture yet, but the simple idea we have is, Congress policies, as part of the enforcement workflow, should set the VM placement constraints, and feed any additional data to be used for scheduling/placement decisions, which will be consumed dynamically by the Solver Scheduler, and after the delegation, the Solver scheduler module will calculate the placement decisions, and complete the VM initial placement or call the VM migration APIs and enable the required migrations. Thanks, Yathi. On 2/12/15, 10:02 AM, Tim Hinrichs thinri...@vmware.commailto:thinri...@vmware.com wrote: Hi Debo and Yathiraj, I took a third look at the
Re: [openstack-dev] [Congress][Delegation] Google doc for working notes
Hi Debo and Yathiraj, I took a third look at the solver-scheduler docs and code with your comments in mind. A few things jumped out. 1) Choice of LP solver. I see solver-scheduler uses Pulp, which was on the Congress short list as well. So we’re highly aligned on the choice of underlying solver. 2) User control over VM-placement. To choose the criteria for VM-placement, the solver-scheduler user picks from a list of predefined options, e.g. ActiveHostConstraint, MaxRamAllocationPerHostConstraint. We’re investigating a slightly different approach, where the user defines the criteria for VM-placement by writing any policy they like in Datalog. Under the hood we then convert that Datalog to an LP problem. From the developer’s perspective, with the Congress approach we don’t attempt to anticipate the different policies the user might want and write code for each policy; instead, we as developers write a translator from Datalog to LP. From the user’s perspective, the difference is that if the option they want isn’t on the solver-scheduler's list, they’re out of luck or need to write the code themselves. But with the Congress approach, they can write any VM-placement policy they like. What I’d like to see is the best of both worlds. Users write Datalog policies describing whatever VM-placement policy they want. If the policy they’ve written is on the solver-scheduler’s list of options, we use the hard-coded implementation, but if the policy isn’t on that list we translate directly to LP. This approach gives us the ability to write custom code to handle common cases while at the same time letting users write whatever policy they like. 3) API and architecture. Today the solver-scheduler's VM-placement policy is defined at config-time (i.e. not run-time). Am I correct that this limitation is only because there’s no API call to set the solver-scheduler’s policy? Or is there some other reason the policy is set at config-time? Congress policies change at runtime, so we’ll definitely need a VM-placement engine whose policy can be changed at run-time as well. If we focus on just migration (and not provisioning), we can build a VM-placement engine that sits outside of Nova that has an API call that allows us to set policy at runtime. We can also set up that engine to get data updates that influence the policy. We were planning on creating this kind of VM-placement engine within Congress as a node on the DSE (our message bus). This is convenient because all nodes on the DSE run in their own thread, any node on the DSE can subscribe to any data from any other node (e.g. ceilometer’s data), and the algorithms for translating Datalog to LP look to be quite similar to the algorithms we’re using in our domain-agnostic policy engine. Tim On Feb 11, 2015, at 4:50 PM, Debojyoti Dutta ddu...@gmail.commailto:ddu...@gmail.com wrote: Hi Tim: moving our thread to the mailer. Excited to collaborate! From: Debo~ Dutta dedu...@cisco.commailto:dedu...@cisco.com Date: Wednesday, February 11, 2015 at 4:48 PM To: Tim Hinrichs thinri...@vmware.commailto:thinri...@vmware.com Cc: Yathiraj Udupi (yudupi) yud...@cisco.commailto:yud...@cisco.com, Gokul B Kandiraju go...@us.ibm.commailto:go...@us.ibm.com, Prabhakar Kudva ku...@us.ibm.commailto:ku...@us.ibm.com, ruby.krishnasw...@orange.commailto:ruby.krishnasw...@orange.com ruby.krishnasw...@orange.commailto:ruby.krishnasw...@orange.com, dilik...@in.ibm.commailto:dilik...@in.ibm.com dilik...@in.ibm.commailto:dilik...@in.ibm.com, Norival Figueira nfigu...@brocade.commailto:nfigu...@brocade.com, Ramki Krishnan r...@brocade.commailto:r...@brocade.com, Xinyuan Huang (xinyuahu) xinyu...@cisco.commailto:xinyu...@cisco.com, Rishabh Jain -X (rishabja - AAP3 INC at Cisco) risha...@cisco.commailto:risha...@cisco.com Subject: Re: Nova solver scheduler and Congress Hi Tim To address your particular questions: 1. translate some policy language into constraints for the LP/CVP and we had left that to congress hoping to integrate when the policy efforts in openstack were ready (our initial effort was pre congress) 2. For migrations: we are currently doing that – its about incremental constraints into the same solver. Hence its a small deal …. Joining forces is a terrific idea. Would love to join the IRC call and see how we can build cool stuff in the community together. I hope we don’t have to replicate the vm placement engine while the work that was done in the community does something very similar (and be adapted) debo From: Tim Hinrichs thinri...@vmware.commailto:thinri...@vmware.com Date: Wednesday, February 11, 2015 at 4:43 PM To: Debo~ Dutta dedu...@cisco.commailto:dedu...@cisco.com Cc: Yathiraj Udupi (yudupi) yud...@cisco.commailto:yud...@cisco.com, Gokul B Kandiraju go...@us.ibm.commailto:go...@us.ibm.com, Prabhakar Kudva ku...@us.ibm.commailto:ku...@us.ibm.com,
[openstack-dev] [Congress][Delegation] Google doc for working notes
Hi all, A (growing) group of folks are interested in working on the problem of delegating policy from Congress to domain-specific policy engines. We started looking at an NFV use case: migrating VMs to reduce energy consumption. In particular we’re looking into building a VM-placement policy engine built on top of a linear programming solver. Here’s a doc with some working notes where we’re trying to figure out how to do the translation from Congress policy to the linear program. https://docs.google.com/document/d/1ksDilJYXV-5AXWON8PLMedDKr9NpS8VbT0jIy_MIEtI/edit?usp=sharing Tim __ 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
Re: [openstack-dev] [Congress][Delegation] Google doc for working notes
Hi Tim: moving our thread to the mailer. Excited to collaborate! From: Debo~ Dutta dedu...@cisco.com Date: Wednesday, February 11, 2015 at 4:48 PM To: Tim Hinrichs thinri...@vmware.com Cc: Yathiraj Udupi (yudupi) yud...@cisco.com, Gokul B Kandiraju go...@us.ibm.com, Prabhakar Kudva ku...@us.ibm.com, ruby.krishnasw...@orange.com ruby.krishnasw...@orange.com, dilik...@in.ibm.com dilik...@in.ibm.com, Norival Figueira nfigu...@brocade.com, Ramki Krishnan r...@brocade.com, Xinyuan Huang (xinyuahu) xinyu...@cisco.com, Rishabh Jain -X (rishabja - AAP3 INC at Cisco) risha...@cisco.com Subject: Re: Nova solver scheduler and Congress Hi Tim To address your particular questions: 1. translate some policy language into constraints for the LP/CVP and we had left that to congress hoping to integrate when the policy efforts in openstack were ready (our initial effort was pre congress) 2. For migrations: we are currently doing that – its about incremental constraints into the same solver. Hence its a small deal …. Joining forces is a terrific idea. Would love to join the IRC call and see how we can build cool stuff in the community together. I hope we don’t have to replicate the vm placement engine while the work that was done in the community does something very similar (and be adapted) debo From: Tim Hinrichs thinri...@vmware.com Date: Wednesday, February 11, 2015 at 4:43 PM To: Debo~ Dutta dedu...@cisco.com Cc: Yathiraj Udupi (yudupi) yud...@cisco.com, Gokul B Kandiraju go...@us.ibm.com, Prabhakar Kudva ku...@us.ibm.com, ruby.krishnasw...@orange.com ruby.krishnasw...@orange.com, dilik...@in.ibm.com dilik...@in.ibm.com, Norival Figueira nfigu...@brocade.com, Ramki Krishnan r...@brocade.com, Xinyuan Huang (xinyuahu) xinyu...@cisco.com, Rishabh Jain -X (rishabja - AAP3 INC at Cisco) risha...@cisco.com Subject: Re: Nova solver scheduler and Congress Hi Debo, The 2 efforts share great similarities, which was why we investigated the state of solver-scheduler. From what I understand, (i) solver-scheduler doesn’t currently have a policy language and (ii) it doesn’t do migrations. (I realize these are both in the works.) We needed both and wanted to make progress before those were complete. In the long run, it may make perfect sense to replace our vm-placement engine with yours. So joining forces sounds like a good idea. At the very *least* we ought to keep up to date with each other’s progress. I’m starting to wonder if we ought to schedule a (bi-) weekly IRC for this topic. Tim On Feb 11, 2015, at 4:35 PM, Debo Dutta (dedutta) dedu...@cisco.com wrote: Hi Tim This looks awesome. Trying to figure out how this approach is different from the solver scheduler effort we did? We would be happy to fold our solver scheduler effort into this (that way you also get code up and running) Will also respond on the thread. thx debo From: Tim Hinrichs thinri...@vmware.com Date: Wednesday, February 11, 2015 at 4:11 PM To: Yathiraj Udupi (yudupi) yud...@cisco.com Cc: Gokul B Kandiraju go...@us.ibm.com, Prabhakar Kudva ku...@us.ibm.com, ruby.krishnasw...@orange.com ruby.krishnasw...@orange.com, dilik...@in.ibm.com dilik...@in.ibm.com, Norival Figueira nfigu...@brocade.com, Ramki Krishnan r...@brocade.com, Xinyuan Huang (xinyuahu) xinyu...@cisco.com, Rishabh Jain -X (rishabja - AAP3 INC at Cisco) risha...@cisco.com, Debo~ Dutta dedu...@cisco.com Subject: Re: Nova solver scheduler and Congress Hi Yathiraj, The group is getting big enough that we’ve decided to move the entire discussion to the openstack-dev mailing list. I sent a note today with the google doc we’re working on. We’re trying to include [Congress][Delegation] in the subject line of relevant posts. Here’s the gdoc. https://docs.google.com/document/d/1ksDilJYXV-5AXWON8PLMedDKr9NpS8VbT0jIy_MIEtI/edit?usp=sharing https://urldefense.proofpoint.com/v2/url?u=https-3A__docs.google.com_document_d_1ksDilJYXV-2D5AXWON8PLMedDKr9NpS8VbT0jIy-5FMIEtI_edit-3Fusp-3Dsharingd=AwMF-gc=Sqcl0Ez6M0X8aeM67LKIiDJAXVeAw-YihVMNtXt-uEsr=B6BWd4kFfgOzAREgThxkmTZKy7dDXE2-eBAmL0PBK7sm=no-emyiErtYa3_zneDNhY78LG0mCHc0bgMpXi1StZ7As=rLY2ACQqD5EQn3MgOnoX8M_zr9254v-FqhF56wfGpice= Tim On Feb 10, 2015, at 11:10 AM, Yathiraj Udupi (yudupi) yud...@cisco.com wrote: Hi Tim, Thanks for your response. I think Congress will have to appreciate different APIs interacting with multiple components in the OpenStack ecosystem. So I will be happy to help figure out the integration plan in general from the Congress side. I will be very interested and glad to participate in the discussions of designing these interfaces in Congress. Please share any preliminary designs you may have. I had participated in one of the Congress mid-cycle meet ups, and I am interested in the upcoming work on these kind of enforcement aspects (reactive, proactive) of Congress. In terms of Nova scheduling via Solver scheduler, it will also help us be ready with the integration points. Let’s be