Re: [openstack-dev] [oslo.messaging] Performance testing. Initial steps.
On Wed, Jan 28, 2015, at 03:23 AM, Denis Makogon wrote: > On Tue, Jan 27, 2015 at 10:26 PM, Gordon Sim wrote: > > > On 01/27/2015 06:31 PM, Doug Hellmann wrote: > > > >> On Tue, Jan 27, 2015, at 12:28 PM, Denis Makogon wrote: > >> > >>> I'd like to build tool that would be able to profile messaging over > >>> various deployments. This "tool" would give me an ability to compare > >>> results of performance testing produced by native tools and > >>> oslo.messaging-based tool, eventually it would lead us into digging into > >>> code and trying to figure out where "bad things" are happening (that's > >>> the > >>> actual place where we would need to profile messaging code). Correct me > >>> if > >>> i'm wrong. > >>> > >> > >> It would be interesting to have recommendations for deployment of rabbit > >> or qpid based on performance testing with oslo.messaging. It would also > >> be interesting to have recommendations for changes to the implementation > >> of oslo.messaging based on performance testing. I'm not sure you want to > >> do full-stack testing for the latter, though. > >> > >> Either way, I think you would be able to start the testing without any > >> changes in oslo.messaging. > >> > > > > I agree. I think the first step is to define what to measure and then > > construct an application using olso.messaging that allows the data of > > interest to be captured using different drivers and indeed different > > configurations of a given driver. > > > > I wrote a very simple test application to test one aspect that I felt was > > important, namely the scalability of the RPC mechanism as you increase the > > number of clients and servers involved. The code I used is > > https://github.com/grs/ombt, its probably stale at the moment, I only > > link to it as an example of approach. > > > > Using that test code I was then able to compare performance in this one > > aspect across drivers (the 'rabbit', 'qpid' and new amqp 1.0 based drivers > > _ I wanted to try zmq, but couldn't figure out how to get it working at the > > time), and for different deployment options using a given driver (amqp 1.0 > > using qpidd or qpid dispatch router in either standalone or with multiple > > connected routers). > > > > There are of course several other aspects that I think would be important > > to explore: notifications, more specific variations in the RPC 'topology' > > i.e. number of clients on given server number of servers in single group > > etc, and a better tool (or set of tools) would allow all of these to be > > explored. > > > > From my experimentation, I believe the biggest differences in scalability > > are going to come not from optimising the code in oslo.messaging so much as > > choosing different patterns for communication. Those choices may be > > constrained by other aspects as well of course, notably approach to > > reliability. > > > > > > > After couple internal discussions and hours of investigations, i think > i've > foung the most applicabale solution > that will accomplish performance testing approach and will eventually be > evaluated as messaging drivers > configuration and AMQP service deployment recommendataion. > > Solution that i've been talking about is already pretty well-known across > OpenStack components - Rally and its scenarios. > Why it would be the best option? Rally scenarios would not touch > messaging > core part. Scenarios are gate-able. > Even if we're talking about internal testing, scenarios are very useful > in > this case, > since they are something that can be tuned/configured taking into account > environment needs. > > Doug, Gordon, what do you think about bringing scenarios into messaging? I think I need more detail about what you mean by that. Doug > > __ > > 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 > > > > > Kind regards, > Denis M. > __ > 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 __ 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] [oslo.messaging] Performance testing. Initial steps.
On Wed, Jan 28, 2015 at 11:39 AM, Flavio Percoco wrote: > On 28/01/15 10:23 +0200, Denis Makogon wrote: > >> >> >> On Tue, Jan 27, 2015 at 10:26 PM, Gordon Sim wrote: >> >>On 01/27/2015 06:31 PM, Doug Hellmann wrote: >> On Tue, Jan 27, 2015, at 12:28 PM, Denis Makogon wrote: >> I'd like to build tool that would be able to profile >> messaging over >>various deployments. This "tool" would give me an ability to >>compare >>results of performance testing produced by native tools and >>oslo.messaging-based tool, eventually it would lead us into >> digging >>into >>code and trying to figure out where "bad things" are happening >>(that's >>the >>actual place where we would need to profile messaging code). >>Correct me >>if >>i'm wrong. >> >> >>It would be interesting to have recommendations for deployment of >>rabbit >>or qpid based on performance testing with oslo.messaging. It would >> also >>be interesting to have recommendations for changes to the >>implementation >>of oslo.messaging based on performance testing. I'm not sure you >> want >>to >>do full-stack testing for the latter, though. >> >>Either way, I think you would be able to start the testing without >> any >>changes in oslo.messaging. >> >>I agree. I think the first step is to define what to measure and then >>construct an application using olso.messaging that allows the data of >>interest to be captured using different drivers and indeed different >>configurations of a given driver. >> >>I wrote a very simple test application to test one aspect that I felt >> was >>important, namely the scalability of the RPC mechanism as you increase >> the >>number of clients and servers involved. The code I used is https:// >>github.com/grs/ombt, its probably stale at the moment, I only link to >> it as >>an example of approach. >> >>Using that test code I was then able to compare performance in this one >>aspect across drivers (the 'rabbit', 'qpid' and new amqp 1.0 based >> drivers >>_ I wanted to try zmq, but couldn't figure out how to get it working >> at the >>time), and for different deployment options using a given driver (amqp >> 1.0 >>using qpidd or qpid dispatch router in either standalone or with >> multiple >>connected routers). >> >>There are of course several other aspects that I think would be >> important >>to explore: notifications, more specific variations in the RPC >> 'topology' >>i.e. number of clients on given server number of servers in single >> group >>etc, and a better tool (or set of tools) would allow all of these to be >>explored. >> >>From my experimentation, I believe the biggest differences in >> scalability >>are going to come not from optimising the code in oslo.messaging so >> much as >>choosing different patterns for communication. Those choices may be >>constrained by other aspects as well of course, notably approach to >>reliability. >> >> >> >> >> After couple internal discussions and hours of investigations, i think >> i've >> foung the most applicabale solution >> that will accomplish performance testing approach and will eventually be >> evaluated as messaging drivers >> configuration and AMQP service deployment recommendataion. >> >> Solution that i've been talking about is already pretty well-known across >> OpenStack components - Rally and its scenarios. >> Why it would be the best option? Rally scenarios would not touch messaging >> core part. Scenarios are gate-able. >> Even if we're talking about internal testing, scenarios are very useful >> in this >> case, >> since they are something that can be tuned/configured taking into account >> environment needs. >> >> Doug, Gordon, what do you think about bringing scenarios into messaging? >> > > I personally wouldn't mind having them but I'd like us to first > discuss what kind of scenarios we want to test. > > I'm assuming these scenarios would be pure oslo.messaging scenarios > and they won't require any of the openstack services. Therefore, I > guess these scenarios would test things like performance with many consumers, performance with several (a)synchronous calls, etc. What > performance means in this context will have to be discussed as well. > Correct, oslo.messaging scenarios would expect to have only AMQP service and nothing else. Yes, that's what i've been thinking about. Also, i'd like to share doc that i've found, see [1]. As i can see it would be more than useful to enable next scenarios: - Single multi-thread publisher (rpc client) against single multi-thread consumer - using RPC cast/call methods try to measure time between request and response. - Multiple multi-thread publisher
Re: [openstack-dev] [oslo.messaging] Performance testing. Initial steps.
On 28/01/15 10:23 +0200, Denis Makogon wrote: On Tue, Jan 27, 2015 at 10:26 PM, Gordon Sim wrote: On 01/27/2015 06:31 PM, Doug Hellmann wrote: On Tue, Jan 27, 2015, at 12:28 PM, Denis Makogon wrote: I'd like to build tool that would be able to profile messaging over various deployments. This "tool" would give me an ability to compare results of performance testing produced by native tools and oslo.messaging-based tool, eventually it would lead us into digging into code and trying to figure out where "bad things" are happening (that's the actual place where we would need to profile messaging code). Correct me if i'm wrong. It would be interesting to have recommendations for deployment of rabbit or qpid based on performance testing with oslo.messaging. It would also be interesting to have recommendations for changes to the implementation of oslo.messaging based on performance testing. I'm not sure you want to do full-stack testing for the latter, though. Either way, I think you would be able to start the testing without any changes in oslo.messaging. I agree. I think the first step is to define what to measure and then construct an application using olso.messaging that allows the data of interest to be captured using different drivers and indeed different configurations of a given driver. I wrote a very simple test application to test one aspect that I felt was important, namely the scalability of the RPC mechanism as you increase the number of clients and servers involved. The code I used is https:// github.com/grs/ombt, its probably stale at the moment, I only link to it as an example of approach. Using that test code I was then able to compare performance in this one aspect across drivers (the 'rabbit', 'qpid' and new amqp 1.0 based drivers _ I wanted to try zmq, but couldn't figure out how to get it working at the time), and for different deployment options using a given driver (amqp 1.0 using qpidd or qpid dispatch router in either standalone or with multiple connected routers). There are of course several other aspects that I think would be important to explore: notifications, more specific variations in the RPC 'topology' i.e. number of clients on given server number of servers in single group etc, and a better tool (or set of tools) would allow all of these to be explored. From my experimentation, I believe the biggest differences in scalability are going to come not from optimising the code in oslo.messaging so much as choosing different patterns for communication. Those choices may be constrained by other aspects as well of course, notably approach to reliability. After couple internal discussions and hours of investigations, i think i've foung the most applicabale solution that will accomplish performance testing approach and will eventually be evaluated as messaging drivers configuration and AMQP service deployment recommendataion. Solution that i've been talking about is already pretty well-known across OpenStack components - Rally and its scenarios. Why it would be the best option? Rally scenarios would not touch messaging core part. Scenarios are gate-able. Even if we're talking about internal testing, scenarios are very useful in this case, since they are something that can be tuned/configured taking into account environment needs. Doug, Gordon, what do you think about bringing scenarios into messaging? I personally wouldn't mind having them but I'd like us to first discuss what kind of scenarios we want to test. I'm assuming these scenarios would be pure oslo.messaging scenarios and they won't require any of the openstack services. Therefore, I guess these scenarios would test things like performance with many consumers, performance with several (a)synchronous calls, etc. What performance means in this context will have to be discussed as well. In addition to the above, it'd be really interesting if we could have tests for things like reconnects delays, which I think is doable with Rally. Am I right? Cheers, Flavio __ 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 Kind regards, Denis M. __ 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 -- @flaper87 Flavio Percoco pgpnc_PEFy9KP.pgp
Re: [openstack-dev] [oslo.messaging] Performance testing. Initial steps.
On Tue, Jan 27, 2015 at 10:26 PM, Gordon Sim wrote: > On 01/27/2015 06:31 PM, Doug Hellmann wrote: > >> On Tue, Jan 27, 2015, at 12:28 PM, Denis Makogon wrote: >> >>> I'd like to build tool that would be able to profile messaging over >>> various deployments. This "tool" would give me an ability to compare >>> results of performance testing produced by native tools and >>> oslo.messaging-based tool, eventually it would lead us into digging into >>> code and trying to figure out where "bad things" are happening (that's >>> the >>> actual place where we would need to profile messaging code). Correct me >>> if >>> i'm wrong. >>> >> >> It would be interesting to have recommendations for deployment of rabbit >> or qpid based on performance testing with oslo.messaging. It would also >> be interesting to have recommendations for changes to the implementation >> of oslo.messaging based on performance testing. I'm not sure you want to >> do full-stack testing for the latter, though. >> >> Either way, I think you would be able to start the testing without any >> changes in oslo.messaging. >> > > I agree. I think the first step is to define what to measure and then > construct an application using olso.messaging that allows the data of > interest to be captured using different drivers and indeed different > configurations of a given driver. > > I wrote a very simple test application to test one aspect that I felt was > important, namely the scalability of the RPC mechanism as you increase the > number of clients and servers involved. The code I used is > https://github.com/grs/ombt, its probably stale at the moment, I only > link to it as an example of approach. > > Using that test code I was then able to compare performance in this one > aspect across drivers (the 'rabbit', 'qpid' and new amqp 1.0 based drivers > _ I wanted to try zmq, but couldn't figure out how to get it working at the > time), and for different deployment options using a given driver (amqp 1.0 > using qpidd or qpid dispatch router in either standalone or with multiple > connected routers). > > There are of course several other aspects that I think would be important > to explore: notifications, more specific variations in the RPC 'topology' > i.e. number of clients on given server number of servers in single group > etc, and a better tool (or set of tools) would allow all of these to be > explored. > > From my experimentation, I believe the biggest differences in scalability > are going to come not from optimising the code in oslo.messaging so much as > choosing different patterns for communication. Those choices may be > constrained by other aspects as well of course, notably approach to > reliability. > > > After couple internal discussions and hours of investigations, i think i've foung the most applicabale solution that will accomplish performance testing approach and will eventually be evaluated as messaging drivers configuration and AMQP service deployment recommendataion. Solution that i've been talking about is already pretty well-known across OpenStack components - Rally and its scenarios. Why it would be the best option? Rally scenarios would not touch messaging core part. Scenarios are gate-able. Even if we're talking about internal testing, scenarios are very useful in this case, since they are something that can be tuned/configured taking into account environment needs. Doug, Gordon, what do you think about bringing scenarios into messaging? __ > 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 > Kind regards, Denis M. __ 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] [oslo.messaging] Performance testing. Initial steps.
On 01/27/2015 06:31 PM, Doug Hellmann wrote: On Tue, Jan 27, 2015, at 12:28 PM, Denis Makogon wrote: I'd like to build tool that would be able to profile messaging over various deployments. This "tool" would give me an ability to compare results of performance testing produced by native tools and oslo.messaging-based tool, eventually it would lead us into digging into code and trying to figure out where "bad things" are happening (that's the actual place where we would need to profile messaging code). Correct me if i'm wrong. It would be interesting to have recommendations for deployment of rabbit or qpid based on performance testing with oslo.messaging. It would also be interesting to have recommendations for changes to the implementation of oslo.messaging based on performance testing. I'm not sure you want to do full-stack testing for the latter, though. Either way, I think you would be able to start the testing without any changes in oslo.messaging. I agree. I think the first step is to define what to measure and then construct an application using olso.messaging that allows the data of interest to be captured using different drivers and indeed different configurations of a given driver. I wrote a very simple test application to test one aspect that I felt was important, namely the scalability of the RPC mechanism as you increase the number of clients and servers involved. The code I used is https://github.com/grs/ombt, its probably stale at the moment, I only link to it as an example of approach. Using that test code I was then able to compare performance in this one aspect across drivers (the 'rabbit', 'qpid' and new amqp 1.0 based drivers _ I wanted to try zmq, but couldn't figure out how to get it working at the time), and for different deployment options using a given driver (amqp 1.0 using qpidd or qpid dispatch router in either standalone or with multiple connected routers). There are of course several other aspects that I think would be important to explore: notifications, more specific variations in the RPC 'topology' i.e. number of clients on given server number of servers in single group etc, and a better tool (or set of tools) would allow all of these to be explored. From my experimentation, I believe the biggest differences in scalability are going to come not from optimising the code in oslo.messaging so much as choosing different patterns for communication. Those choices may be constrained by other aspects as well of course, notably approach to reliability. __ 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] [oslo.messaging] Performance testing. Initial steps.
On Tue, Jan 27, 2015, at 12:28 PM, Denis Makogon wrote: > On Tue, Jan 27, 2015 at 7:15 PM, Doug Hellmann > wrote: > > > > > > > On Tue, Jan 27, 2015, at 10:56 AM, Denis Makogon wrote: > > > On Thu, Jan 15, 2015 at 8:56 PM, Doug Hellmann > > > wrote: > > > > > > > > > > > > On Jan 15, 2015, at 1:30 PM, Denis Makogon > > > > wrote: > > > > > > > > > > Good day to All, > > > > > > > > > > The question that i’d like to raise here is not simple one, so i’d > > like > > > > to involve as much readers as i can. I’d like to speak about > > oslo.messaging > > > > performance testing. As community we’ve put lots of efforts in making > > > > oslo.messaging widely used drivers stable as much as possible. > > Stability is > > > > a good thing, but is it enough for saying “works well”? I’d say that > > it’s > > > > not. > > > > > Since oslo.messaging uses driver-based messaging workflow, it makes > > > > sense to dig into each driver and collect all required/possible > > performance > > > > metrics. > > > > > First of all, it does make sense to figure out how to perform > > > > performance testing, first that came into my mind is to simulate high > > load > > > > on one of corresponding drivers. Here comes the question of how it can > > be > > > > accomplished withing available oslo.messaging tools - high load on any > > > > driver can perform an application that: > > > > > • can populate multiple emitters(rpc clients) and consumers > > (rpc > > > > servers). > > > > > • can force clients to send messages of pre-defined number of > > > > messages of any length. > > > > > > > > That makes sense. > > > > > > > > > Another thing is why do we need such thing. Profiling, performance > > > > testing can improve the way in which our drivers were implemented. It > > can > > > > show us actual “bottlenecks” in messaging process, in general. In some > > > > cases it does make sense to figure out where problem takes its place - > > > > whether AMQP causes messaging problems or certain driver that speaks to > > > > AMQP fails. > > > > > Next thing that i want to discuss the architecture of > > > > profiling/performance testing. As i can see it seemed to be a “good” > > way to > > > > add profiling code to each driver. If there’s any objection or better > > > > solution, please bring them to the light. > > > > > > > > What sort of extra profiling code do you anticipate needing? > > > > > > > > > > > As i can foresee (taking into account [1]) couple decorators, possibly > > > one > > > that handles metering process. The biggest part of code will take > > > highload > > > tool that'll be a part of messaging. But another question adding certain > > > dependecies to the project. > > > > > > > > > > > Once we’d have final design for profiling we would need to figure out > > > > tools for profiling. After searching over the web, i found pretty > > > > interesting topic related to python profiling [1]. After certain > > > > investigations it does makes sense discuss next profiling > > options(apply one > > > > or both): > > > > > • Line-by-line timing and execution frequency with a profiler > > > > (there are possible Pros and Cons, but i would say the per-line > > statistics > > > > is more than appreciable at initial performance testing steps) > > > > > • Memory/CPU consumption > > > > > Metrics. The most useful metric for us is time, any time-based > > metric, > > > > since it is very useful to know at which step or/and by whom > > delay/timeout > > > > caused, for example, so as it said, we would be able to figure out > > whether > > > > AMQP or driver fails to do what it was designed for. > > > > > Before proposing spec i’d like to figure out any other requirements, > > use > > > > cases and restrictions for messaging performance testing. Also, if > > there > > > > any stories of success in boosting python performance - feel free to > > share > > > > it. > > > > > > > > The metrics to measure depend on the goal. Do we think the messaging > > code > > > > is using too much memory? Is it too slow? Or is there something else > > > > causing concern? > > > > > > > > It does make sense to have profiling for cases when trying to upscale > > > cluster and it'll be a good thing to have an ability to figure out if > > > scaled AMQP service has it's best configuration (i guess here would come > > > the question about doing performance testing using well-known tools), and > > > the most interesting question is about how messaging driver decreases (or > > > leaves untouched) throughput between RPC client and server. This metering > > > results can be compared to those tools that were designed for performance > > > testing. And that's why it'll be good step forward having > > > profiling/performance testing using high load technic. > > > > That makes it sound like you want to build performance testing tools for > > the infrastructure oslo.messaging is using, and not for oslo.messaging > > itself. Is that right? > > >
Re: [openstack-dev] [oslo.messaging] Performance testing. Initial steps.
On Tue, Jan 27, 2015 at 7:15 PM, Doug Hellmann wrote: > > > On Tue, Jan 27, 2015, at 10:56 AM, Denis Makogon wrote: > > On Thu, Jan 15, 2015 at 8:56 PM, Doug Hellmann > > wrote: > > > > > > > > > On Jan 15, 2015, at 1:30 PM, Denis Makogon > > > wrote: > > > > > > > > Good day to All, > > > > > > > > The question that i’d like to raise here is not simple one, so i’d > like > > > to involve as much readers as i can. I’d like to speak about > oslo.messaging > > > performance testing. As community we’ve put lots of efforts in making > > > oslo.messaging widely used drivers stable as much as possible. > Stability is > > > a good thing, but is it enough for saying “works well”? I’d say that > it’s > > > not. > > > > Since oslo.messaging uses driver-based messaging workflow, it makes > > > sense to dig into each driver and collect all required/possible > performance > > > metrics. > > > > First of all, it does make sense to figure out how to perform > > > performance testing, first that came into my mind is to simulate high > load > > > on one of corresponding drivers. Here comes the question of how it can > be > > > accomplished withing available oslo.messaging tools - high load on any > > > driver can perform an application that: > > > > • can populate multiple emitters(rpc clients) and consumers > (rpc > > > servers). > > > > • can force clients to send messages of pre-defined number of > > > messages of any length. > > > > > > That makes sense. > > > > > > > Another thing is why do we need such thing. Profiling, performance > > > testing can improve the way in which our drivers were implemented. It > can > > > show us actual “bottlenecks” in messaging process, in general. In some > > > cases it does make sense to figure out where problem takes its place - > > > whether AMQP causes messaging problems or certain driver that speaks to > > > AMQP fails. > > > > Next thing that i want to discuss the architecture of > > > profiling/performance testing. As i can see it seemed to be a “good” > way to > > > add profiling code to each driver. If there’s any objection or better > > > solution, please bring them to the light. > > > > > > What sort of extra profiling code do you anticipate needing? > > > > > > > > As i can foresee (taking into account [1]) couple decorators, possibly > > one > > that handles metering process. The biggest part of code will take > > highload > > tool that'll be a part of messaging. But another question adding certain > > dependecies to the project. > > > > > > > > Once we’d have final design for profiling we would need to figure out > > > tools for profiling. After searching over the web, i found pretty > > > interesting topic related to python profiling [1]. After certain > > > investigations it does makes sense discuss next profiling > options(apply one > > > or both): > > > > • Line-by-line timing and execution frequency with a profiler > > > (there are possible Pros and Cons, but i would say the per-line > statistics > > > is more than appreciable at initial performance testing steps) > > > > • Memory/CPU consumption > > > > Metrics. The most useful metric for us is time, any time-based > metric, > > > since it is very useful to know at which step or/and by whom > delay/timeout > > > caused, for example, so as it said, we would be able to figure out > whether > > > AMQP or driver fails to do what it was designed for. > > > > Before proposing spec i’d like to figure out any other requirements, > use > > > cases and restrictions for messaging performance testing. Also, if > there > > > any stories of success in boosting python performance - feel free to > share > > > it. > > > > > > The metrics to measure depend on the goal. Do we think the messaging > code > > > is using too much memory? Is it too slow? Or is there something else > > > causing concern? > > > > > > It does make sense to have profiling for cases when trying to upscale > > cluster and it'll be a good thing to have an ability to figure out if > > scaled AMQP service has it's best configuration (i guess here would come > > the question about doing performance testing using well-known tools), and > > the most interesting question is about how messaging driver decreases (or > > leaves untouched) throughput between RPC client and server. This metering > > results can be compared to those tools that were designed for performance > > testing. And that's why it'll be good step forward having > > profiling/performance testing using high load technic. > > That makes it sound like you want to build performance testing tools for > the infrastructure oslo.messaging is using, and not for oslo.messaging > itself. Is that right? > > I'd like to build tool that would be able to profile messaging over various deployments. This "tool" would give me an ability to compare results of performance testing produced by native tools and oslo.messaging-based tool, eventually it would lead us into digging into code and tryin
Re: [openstack-dev] [oslo.messaging] Performance testing. Initial steps.
On Tue, Jan 27, 2015, at 10:56 AM, Denis Makogon wrote: > On Thu, Jan 15, 2015 at 8:56 PM, Doug Hellmann > wrote: > > > > > > On Jan 15, 2015, at 1:30 PM, Denis Makogon > > wrote: > > > > > > Good day to All, > > > > > > The question that i’d like to raise here is not simple one, so i’d like > > to involve as much readers as i can. I’d like to speak about oslo.messaging > > performance testing. As community we’ve put lots of efforts in making > > oslo.messaging widely used drivers stable as much as possible. Stability is > > a good thing, but is it enough for saying “works well”? I’d say that it’s > > not. > > > Since oslo.messaging uses driver-based messaging workflow, it makes > > sense to dig into each driver and collect all required/possible performance > > metrics. > > > First of all, it does make sense to figure out how to perform > > performance testing, first that came into my mind is to simulate high load > > on one of corresponding drivers. Here comes the question of how it can be > > accomplished withing available oslo.messaging tools - high load on any > > driver can perform an application that: > > > • can populate multiple emitters(rpc clients) and consumers (rpc > > servers). > > > • can force clients to send messages of pre-defined number of > > messages of any length. > > > > That makes sense. > > > > > Another thing is why do we need such thing. Profiling, performance > > testing can improve the way in which our drivers were implemented. It can > > show us actual “bottlenecks” in messaging process, in general. In some > > cases it does make sense to figure out where problem takes its place - > > whether AMQP causes messaging problems or certain driver that speaks to > > AMQP fails. > > > Next thing that i want to discuss the architecture of > > profiling/performance testing. As i can see it seemed to be a “good” way to > > add profiling code to each driver. If there’s any objection or better > > solution, please bring them to the light. > > > > What sort of extra profiling code do you anticipate needing? > > > > > As i can foresee (taking into account [1]) couple decorators, possibly > one > that handles metering process. The biggest part of code will take > highload > tool that'll be a part of messaging. But another question adding certain > dependecies to the project. > > > > > Once we’d have final design for profiling we would need to figure out > > tools for profiling. After searching over the web, i found pretty > > interesting topic related to python profiling [1]. After certain > > investigations it does makes sense discuss next profiling options(apply one > > or both): > > > • Line-by-line timing and execution frequency with a profiler > > (there are possible Pros and Cons, but i would say the per-line statistics > > is more than appreciable at initial performance testing steps) > > > • Memory/CPU consumption > > > Metrics. The most useful metric for us is time, any time-based metric, > > since it is very useful to know at which step or/and by whom delay/timeout > > caused, for example, so as it said, we would be able to figure out whether > > AMQP or driver fails to do what it was designed for. > > > Before proposing spec i’d like to figure out any other requirements, use > > cases and restrictions for messaging performance testing. Also, if there > > any stories of success in boosting python performance - feel free to share > > it. > > > > The metrics to measure depend on the goal. Do we think the messaging code > > is using too much memory? Is it too slow? Or is there something else > > causing concern? > > > > It does make sense to have profiling for cases when trying to upscale > cluster and it'll be a good thing to have an ability to figure out if > scaled AMQP service has it's best configuration (i guess here would come > the question about doing performance testing using well-known tools), and > the most interesting question is about how messaging driver decreases (or > leaves untouched) throughput between RPC client and server. This metering > results can be compared to those tools that were designed for performance > testing. And that's why it'll be good step forward having > profiling/performance testing using high load technic. That makes it sound like you want to build performance testing tools for the infrastructure oslo.messaging is using, and not for oslo.messaging itself. Is that right? Doug > > > > > > > > > > > > > > [1] http://www.huyng.com/posts/python-performance-analysis/ > > > > > > Kind regards, > > > Denis Makogon > > > IRC: denis_makogon > > > dmako...@mirantis.com > > > > > > > > __ > > > 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] [oslo.messaging] Performance testing. Initial steps.
On Thu, Jan 15, 2015 at 8:56 PM, Doug Hellmann wrote: > > > On Jan 15, 2015, at 1:30 PM, Denis Makogon > wrote: > > > > Good day to All, > > > > The question that i’d like to raise here is not simple one, so i’d like > to involve as much readers as i can. I’d like to speak about oslo.messaging > performance testing. As community we’ve put lots of efforts in making > oslo.messaging widely used drivers stable as much as possible. Stability is > a good thing, but is it enough for saying “works well”? I’d say that it’s > not. > > Since oslo.messaging uses driver-based messaging workflow, it makes > sense to dig into each driver and collect all required/possible performance > metrics. > > First of all, it does make sense to figure out how to perform > performance testing, first that came into my mind is to simulate high load > on one of corresponding drivers. Here comes the question of how it can be > accomplished withing available oslo.messaging tools - high load on any > driver can perform an application that: > > • can populate multiple emitters(rpc clients) and consumers (rpc > servers). > > • can force clients to send messages of pre-defined number of > messages of any length. > > That makes sense. > > > Another thing is why do we need such thing. Profiling, performance > testing can improve the way in which our drivers were implemented. It can > show us actual “bottlenecks” in messaging process, in general. In some > cases it does make sense to figure out where problem takes its place - > whether AMQP causes messaging problems or certain driver that speaks to > AMQP fails. > > Next thing that i want to discuss the architecture of > profiling/performance testing. As i can see it seemed to be a “good” way to > add profiling code to each driver. If there’s any objection or better > solution, please bring them to the light. > > What sort of extra profiling code do you anticipate needing? > > As i can foresee (taking into account [1]) couple decorators, possibly one that handles metering process. The biggest part of code will take highload tool that'll be a part of messaging. But another question adding certain dependecies to the project. > > Once we’d have final design for profiling we would need to figure out > tools for profiling. After searching over the web, i found pretty > interesting topic related to python profiling [1]. After certain > investigations it does makes sense discuss next profiling options(apply one > or both): > > • Line-by-line timing and execution frequency with a profiler > (there are possible Pros and Cons, but i would say the per-line statistics > is more than appreciable at initial performance testing steps) > > • Memory/CPU consumption > > Metrics. The most useful metric for us is time, any time-based metric, > since it is very useful to know at which step or/and by whom delay/timeout > caused, for example, so as it said, we would be able to figure out whether > AMQP or driver fails to do what it was designed for. > > Before proposing spec i’d like to figure out any other requirements, use > cases and restrictions for messaging performance testing. Also, if there > any stories of success in boosting python performance - feel free to share > it. > > The metrics to measure depend on the goal. Do we think the messaging code > is using too much memory? Is it too slow? Or is there something else > causing concern? > > It does make sense to have profiling for cases when trying to upscale cluster and it'll be a good thing to have an ability to figure out if scaled AMQP service has it's best configuration (i guess here would come the question about doing performance testing using well-known tools), and the most interesting question is about how messaging driver decreases (or leaves untouched) throughput between RPC client and server. This metering results can be compared to those tools that were designed for performance testing. And that's why it'll be good step forward having profiling/performance testing using high load technic. > > > > > > > > [1] http://www.huyng.com/posts/python-performance-analysis/ > > > > Kind regards, > > Denis Makogon > > IRC: denis_makogon > > dmako...@mirantis.com > > > > > __ > > 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 > > > __ > 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 > Kind regards, Denis M. __ OpenStack Development Mailing List (not for usage questions) Unsubscribe: openstack-dev-re
Re: [openstack-dev] [oslo.messaging] Performance testing. Initial steps.
> On Jan 15, 2015, at 1:30 PM, Denis Makogon wrote: > > Good day to All, > > The question that i’d like to raise here is not simple one, so i’d like to > involve as much readers as i can. I’d like to speak about oslo.messaging > performance testing. As community we’ve put lots of efforts in making > oslo.messaging widely used drivers stable as much as possible. Stability is a > good thing, but is it enough for saying “works well”? I’d say that it’s not. > Since oslo.messaging uses driver-based messaging workflow, it makes sense to > dig into each driver and collect all required/possible performance metrics. > First of all, it does make sense to figure out how to perform performance > testing, first that came into my mind is to simulate high load on one of > corresponding drivers. Here comes the question of how it can be accomplished > withing available oslo.messaging tools - high load on any driver can perform > an application that: > • can populate multiple emitters(rpc clients) and consumers (rpc > servers). > • can force clients to send messages of pre-defined number of messages > of any length. That makes sense. > Another thing is why do we need such thing. Profiling, performance testing > can improve the way in which our drivers were implemented. It can show us > actual “bottlenecks” in messaging process, in general. In some cases it does > make sense to figure out where problem takes its place - whether AMQP causes > messaging problems or certain driver that speaks to AMQP fails. > Next thing that i want to discuss the architecture of profiling/performance > testing. As i can see it seemed to be a “good” way to add profiling code to > each driver. If there’s any objection or better solution, please bring them > to the light. What sort of extra profiling code do you anticipate needing? > Once we’d have final design for profiling we would need to figure out tools > for profiling. After searching over the web, i found pretty interesting topic > related to python profiling [1]. After certain investigations it does makes > sense discuss next profiling options(apply one or both): > • Line-by-line timing and execution frequency with a profiler (there > are possible Pros and Cons, but i would say the per-line statistics is more > than appreciable at initial performance testing steps) > • Memory/CPU consumption > Metrics. The most useful metric for us is time, any time-based metric, since > it is very useful to know at which step or/and by whom delay/timeout caused, > for example, so as it said, we would be able to figure out whether AMQP or > driver fails to do what it was designed for. > Before proposing spec i’d like to figure out any other requirements, use > cases and restrictions for messaging performance testing. Also, if there any > stories of success in boosting python performance - feel free to share it. The metrics to measure depend on the goal. Do we think the messaging code is using too much memory? Is it too slow? Or is there something else causing concern? > > > > [1] http://www.huyng.com/posts/python-performance-analysis/ > > Kind regards, > Denis Makogon > IRC: denis_makogon > dmako...@mirantis.com > > __ > 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 __ 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
[openstack-dev] [oslo.messaging] Performance testing. Initial steps.
Good day to All, The question that i’d like to raise here is not simple one, so i’d like to involve as much readers as i can. I’d like to speak about oslo.messaging performance testing. As community we’ve put lots of efforts in making oslo.messaging widely used drivers stable as much as possible. Stability is a good thing, but is it enough for saying “works well”? I’d say that it’s not. Since oslo.messaging uses driver-based messaging workflow, it makes sense to dig into each driver and collect all required/possible performance metrics. First of all, it does make sense to figure out how to perform performance testing, first that came into my mind is to simulate high load on one of corresponding drivers. Here comes the question of how it can be accomplished withing available oslo.messaging tools - high load on any driver can perform an application that: - can populate multiple emitters(rpc clients) and consumers (rpc servers). - can force clients to send messages of pre-defined number of messages of any length. Another thing is why do we need such thing. Profiling, performance testing can improve the way in which our drivers were implemented. It can show us actual “bottlenecks” in messaging process, in general. In some cases it does make sense to figure out where problem takes its place - whether AMQP causes messaging problems or certain driver that speaks to AMQP fails. Next thing that i want to discuss the architecture of profiling/performance testing. As i can see it seemed to be a “good” way to add profiling code to each driver. If there’s any objection or better solution, please bring them to the light. Once we’d have final design for profiling we would need to figure out tools for profiling. After searching over the web, i found pretty interesting topic related to python profiling [1]. After certain investigations it does makes sense discuss next profiling options(apply one or both): - Line-by-line timing and execution frequency with a profiler (there are possible Pros and Cons, but i would say the per-line statistics is more than appreciable at initial performance testing steps) - Memory/CPU consumption Metrics. The most useful metric for us is time, any time-based metric, since it is very useful to know at which step or/and by whom delay/timeout caused, for example, so as it said, we would be able to figure out whether AMQP or driver fails to do what it was designed for. Before proposing spec i’d like to figure out any other requirements, use cases and restrictions for messaging performance testing. Also, if there any stories of success in boosting python performance - feel free to share it. [1] http://www.huyng.com/posts/python-performance-analysis/ Kind regards, Denis Makogon IRC: denis_makogon dmako...@mirantis.com __ 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