Re: [Discuss--A proposal for building an application in mahout to measure runtime performance of algorithms in mahout]

[Saikat Kanjilal <sx...@hotmail.com>]

AndrewM/AndrewP and others that may be interested,

I am going to move ahead with the next additions to the runtime performane measurement framework as I have not heard back from either of you on my previous requests, this includes:

1) Organizing the code a bit more so that the driver collects the data into csv files into a pre created directory

2) Adding the zeppelin interface so that one can view the pre and post effects of making changes to a set of hyperparameters for each of the algorithms



Please respond an/or look at my github pull request, I will assume that you guys are in general agreement in the direction that this is headed if I don't hear back otherwise.  Also please do share your thoughts if you see additional things I should think about putting in.



Thanks in advance.


________________________________
From: Saikat Kanjilal <sx...@hotmail.com>
Sent: Wednesday, July 20, 2016 7:55 PM
To: dev@mahout.apache.org
Subject: Re: [Discuss--A proposal for building an application in mahout to measure runtime performance of algorithms in mahout]

Hello Mahout Committers,
I would like to follow up on this and am eager to get traction/feedback on my pull request so far.  Would love to get an initial implementation working for this in the next month so look forward to hearing from folks who would be interested in using this.

Thanks in advance.

Sent from my iPhone

> On Jul 12, 2016, at 10:56 PM, Saikat Kanjilal <sx...@hotmail.com> wrote:
>
> Andrew,
> Ping on this , let me know your thoughts on my pull request.
> Thanks
>
> Sent from my iPad
>
>> On Jul 7, 2016, at 8:15 AM, Andrew Musselman <an...@gmail.com> wrote:
>>
>> Excellent, thanks Saikat; I'll be able to take a look over the weekend.
>>
>>> On Wed, Jul 6, 2016 at 9:37 PM, Saikat Kanjilal <sx...@hotmail.com> wrote:
>>>
>>> Ok folks I've created a pull request here for a barebones runtime
>>> performance measurement framework that:
>>>
>>> 1) measures two simple timings from ssvd and spca
>>>
>>> 2) dumps these timings into a csv file
>>>
>>>
>>> https://github.com/apache/mahout/pull/245
[https://avatars0.githubusercontent.com/u/674374?v=3&s=400]<https://github.com/apache/mahout/pull/245>

Mahout 1869 by skanjila · Pull Request #245 · apache/mahout<https://github.com/apache/mahout/pull/245>
github.com
Added the ability to dump output to csv file



>>>
>>> [https://avatars0.githubusercontent.com/u/674374?v=3&s=400]<
>>> https://github.com/apache/mahout/pull/245>
>>>
>>> Mahout 1869 by skanjila · Pull Request #245 · apache/mahout<
>>> https://github.com/apache/mahout/pull/245>
>>> github.com
>>> Added the ability to dump output to csv file
>>>
>>>
>>>
>>>
>>>
>>> I'd be greatly appreciative if I can get some early feedback on design
>>> before moving forward and making too many more changes and not getting
>>> something included.  I will move ahead with the zeppelin integration in a
>>> few days and reorganize the code a bit to include all the perf related
>>> pieces into one class or trait.
>>>
>>>
>>> Thanks in advance for your help.
>>>
>>>
>>> Thanks in advance.
>>>
>>>
>>> ________________________________
>>> From: Saikat Kanjilal <sx...@hotmail.com>
>>> Sent: Tuesday, June 21, 2016 9:21 PM
>>> To: dev@mahout.apache.org
>>> Subject: RE: [Discuss--A proposal for building an application in mahout to
>>> measure runtime performance of algorithms in mahout]
>>>
>>> Ok, so for now I am able to get around the issues bwlow by working on code
>>> to measure performance times  not requiring the notion of a
>>> DIstributedContext to get this up and running, I have two methods that I am
>>> measuring performance times for,ssvd and spca.   Github repo is here:
>>> https://github.com/skanjila/mahout/tree/mahout-1869
>>> [https://avatars0.githubusercontent.com/u/674374?v=3&s=400]<
>>> https://github.com/skanjila/mahout/tree/mahout-1869>
>>>
>>> skanjila/mahout<https://github.com/skanjila/mahout/tree/mahout-1869>
>>> github.com
>>> mahout - Mirror of Apache Mahout
>>>
>>>
>>>
>>> Please provide feedback as I will now restructure/reorganize code to add
>>> more methods and start work on a perf harness that spits out a report in
>>> csv and then eventually tie this to zeppelin.
>>> I've kept JIRA up to date as well.
>>> Thanks in advance.
>>>
>>>> From: sxk1969@hotmail.com
>>>> To: dev@mahout.apache.org
>>>> Subject: RE: [Discuss--A proposal for building an application in mahout
>>> to measure runtime performance of algorithms in mahout]
>>>> Date: Mon, 20 Jun 2016 20:37:31 -0700
>>>>
>>>> AndrewP et al,Any chance I can get some pointers on the items below,
>>> would love some direction on this.Thanks
>>>>
>>>>> From: sxk1969@hotmail.com
>>>>> To: dev@mahout.apache.org
>>>>> Subject: RE: [Discuss--A proposal for building an application in
>>> mahout to measure runtime performance of algorithms in mahout]
>>>>> Date: Sun, 12 Jun 2016 12:40:26 -0700
>>>>>
>>>>> Hi Folks,I need some input/help here to get me unblocked and moving:
>>>>> 1) I need to reuse/extend the DistributedContext inside the runtime
>>> perf measurement module as all algorithms inside math-scala need this, I
>>> was trying to mimic some of the H2O code and saw that they had their own
>>> engine, I am wondering what the best way is to extend DistributedContext
>>> and get the benefit of an already existing engine without needing to tie
>>> into h2o or flink, or is the only way to add an engine to point to one of
>>> those back ends, ideally I want to build the runtime perf module in a
>>> backend agnostic way and currently I dont see a way around this,
>>> thoughts?2) I also tried to reuse some of the logic inside math-scala but
>>> in digging into this code it seems that this code is strongly tied to scala
>>> test utilities
>>>>>
>>>>> Net-Net: I just need access to the DistributedContext without linking
>>> in any test utilities or backends.
>>>>> Would love some advice on ways to move forward to maximize
>>> reuse.Thanks in advance.
>>>>>
>>>>>> From: sxk1969@hotmail.com
>>>>>> To: dev@mahout.apache.org
>>>>>> Subject: RE: [Discuss--A proposal for building an application in
>>> mahout to measure runtime performance of algorithms in mahout]
>>>>>> Date: Thu, 9 Jun 2016 21:45:13 -0700
>>>>>>
>>>>>> Andrew et al,So I've finally been able to over the past few days got
>>> a self contained module compiling that leverages the DistributedContext,
>>> for starters I copied the NaiveBayes test code, ripped out the test
>>> infrastructure code around it and then added some timers, next steps will
>>> be to dump to csv and eventually to zeppelin, some questions before I get
>>> too far ahead:
>>>>>> 1) I made the design decision to create my own trait and encapsulate
>>> the context within that, I am wondering if I should instead leverage the
>>> context that is already defined in math-scala ,, this however brings its
>>> own complications in that it brings in the MahoutSuite which I'm not sure I
>>> really need, thoughts on this
>>>>>> 2) I need some infrastructure to run the perf framework , I can use
>>> an azure ubuntu vm for now but is there an AWS instance or some other vm I
>>> can eventually use, I would really like to avoid using my mac laptop as a
>>> runtime perf testing environment
>>>>>>
>>>>>> Thanks, I'll update JIRA as I make more headway.
>>>>>>
>>>>>>> From: sxk1969@hotmail.com
>>>>>>> To: dev@mahout.apache.org
>>>>>>> Subject: RE: [Discuss--A proposal for building an application in
>>> mahout to measure runtime performance of algorithms in mahout]
>>>>>>> Date: Mon, 6 Jun 2016 08:58:49 -0700
>>>>>>>
>>>>>>> Andrew,Thanks for the input, I will shift gears a bit and just get
>>> some lightweight code going that calls into mahout algorithms and does a
>>> csv dump out.  Note that I think akka could be a good use for this as you
>>> could make an async call and get back a notification when the csv dump is
>>> finished.  Also I am indeed not focusing on mapreduce algorithms and will
>>> be tackling the algorithms in the math-scala library.  What do you think of
>>> making this a lightweight web based workbench using spray that committers
>>> can run outside of mahout through curl or something, this was my initial
>>> vision in using spray and its good that I'm getting early feedback.
>>>>>>>
>>>>>>> On zeppelin do you think its worthwhile that I incorporate
>>> Trevor's efforts to take that csv and turn that into one or two
>>> visualizations.  I'm trying to understand how that effort may(or may not)
>>> intersect with what I'm trying to accomplish.
>>>>>>> Also point taken on the small data sets.
>>>>>>> Thanks
>>>>>>>
>>>>>>>> From: ap.dev@outlook.com
>>>>>>>> To: dev@mahout.apache.org
>>>>>>>> Subject: Re: [Discuss--A proposal for building an application in
>>> mahout to measure runtime performance of algorithms in mahout]
>>>>>>>> Date: Mon, 6 Jun 2016 15:50:16 +0000
>>>>>>>>
>>>>>>>> Saikat,
>>>>>>>>
>>>>>>>> If you're going to pursue this there is a few things that I
>>> would suggest.  First, keep it light weight.  We don't want to bring a a
>>> lot of extra dependencies or data into the distribution.  I'm not sure what
>>> this means as far as spray/akka, but those seem like overkill in my
>>> opinion. This should be able to be kept down to a simple csv dump I think.
>>>>>>>>
>>>>>>>> Second, use Data that can be either randomly generated with a
>>> seeded RNG, or a function like Mackey-Glass or downloaded (probably best),
>>> and only use a small very small sample in the tests- since they're pretty
>>> long currently. The main point being that we don't want to ship any large
>>> test datasets with the distro.
>>>>>>>>
>>>>>>>> Third, we're not using MapReduce anymore, so focus on algorithms
>>> in the math-scala library (eg. dssvd, thinqr, dals, etc.) as well as Matrix
>>> algebra operations.  That is where i see this being useful, so that we may
>>> compare changes and optimizations going forward.
>>>>>>>>
>>>>>>>> Thanks,
>>>>>>>>
>>>>>>>> Andy
>>>>>>>>
>>>>>>>> ________________________________________
>>>>>>>> From: Saikat Kanjilal <sx...@hotmail.com>
>>>>>>>> Sent: Friday, June 3, 2016 12:35:54 AM
>>>>>>>> To: dev@mahout.apache.org
>>>>>>>> Subject: RE: [Discuss--A proposal for building an application in
>>> mahout to measure runtime performance of algorithms in mahout]
>>>>>>>>
>>>>>>>> Hi All,Created a JIRA ticket and have moved the discussion for
>>> the runtime performance framework  there:
>>>>>>>> https://issues.apache.org/jira/browse/MAHOUT-1869
>>>>>>>> @AndrewP & Trevor I would like to integrate zeppelin into the
>>> runtime performance measurement framework to output some measurement
>>> related data for some of the algorithms.
>>>>>>>> Should I wait till the zeppelin integration is completely
>>> working before I incorporate this piece?
>>>>>>>> Also would really some feedback either on the JIRA ticket or in
>>> response to this thread.Regards
>>>>>>>>
>>>>>>>>> From: sxk1969@hotmail.com
>>>>>>>>> To: dev@mahout.apache.org
>>>>>>>>> Subject: [Discuss--A proposal for building an application in
>>> mahout to measure runtime performance of algorithms in mahout]
>>>>>>>>> Date: Thu, 19 May 2016 21:31:05 -0700
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> This proposal will outline a runtime performance module used
>>> to measure the performance of various algorithms in mahout in the three
>>> major areas, clustering, regression and classification.  The module will be
>>> a spray/scala/akka application which will be run by any current or new
>>> algorithm in mahout and will display a csv file and a set of zeppelin plots
>>> outlining the various criteria for performance.    The goal of releasing
>>> any new build in mahout will be to run a set of tests for each of the
>>> algorithms to compare and contrast some benchmarks from one release to
>>> another.
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> Architecture
>>>>>>>>> The run time performance application will run on top of
>>> spray/scala and akka and will make async api calls into the various mahout
>>> algorithms to generate a cvs file containing data representing the run time
>>> performance measurement calculations for each algorithm of interest as well
>>> as a set of zeppelin plots for displaying some of these results.  The spray
>>> scala architecture will leverage the zeppelin server to create the
>>> visualizations.  The discussion below centers around two types of
>>> algorithms to be addressed by the application.
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> Clustering
>>>>>>>>> The application will consist of a set of rest APIs to do the
>>> following:
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> a) A method to load and execute the run time perf module and
>>> takes as inputs the name of the algorithm (kmeans, fuzzy kmeans) and a
>>> location of a set of files containing various sizes of data sets
>>> /algorithm=clustering/fileLocation=/path/to/files/of/different/datasets/clusters=12,20,30,40
>>> and finally a set of values for the number of clusters to use for each of
>>> the different sizes of the datasets
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> The above API call will return a runId which the client
>>> program can then use to monitor the module
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> b) A method to monitor the application to ensure that its
>>> making progress towards generating the zeppelin plots
>>>>>>>>> /monitor/runId=456
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> The above method will execute asynchronously by calling into
>>> the mahout kmeans (fuzzy kmeans) clustering implementations and will
>>> generate zeppelin plots showing the normalized time on the y axis and the
>>> number of clusters in the x axis.  The spray/scala akka framework will
>>> allow the client application to receive a callback when the run time
>>> performance calculations are actually completed.  For now the calculations
>>> for measuring run time performance will contain: a) the ratio of the number
>>> of points clustered correctly to the total number of points b) the total
>>> time taken for the algorithm to run .  These items will be represented in
>>> separate zeppelin plots.
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> Regression
>>>>>>>>> a) The runtime performance module will run the likelihood
>>> ratio test with a different set of features in every run .  We will
>>> introduce a rest API to run the likelihood ratio test and return the
>>> results, this will once again be an sync call through the spray/akka stack.
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> b) The run time performance module will contain the following
>>> metrics for every algorithm: 1) cpu usage 2) memory usage 3) time taken for
>>> algorithm to converge and run to completion.  These metrics will be
>>> reported on top of the zeppelin graphs for both the regression and the
>>> different clustering algorithms mentioned above.
>>>>>>>>>
>>>>>>>>> How does the application get runThe run time performance
>>> measuring application will get invoked from the command line, eventually it
>>> would be worthwhile to hook this into some sort of integration test suite
>>> to certify the different mahout releases.
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> I will add more thoughts around this and create a JIRA ticket
>>> only once there's enough consensus between the committers that this is
>>> headed in the right direction.  I will also add some more thoughts on
>>> measuring run time performance of some of the other algorithms after some
>>> more research.
>>>>>>>>> Would love feedback or additional things to consider that I
>>> might have missed.  If its more appropriate I can move the discussion to a
>>> jira ticket as well so please let me know.Thanks in advance.
>>>
>>>