airavata-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Pierce, Marlon" <marpi...@iu.edu>
Subject Re: Airavata Registry Considerations
Date Sun, 07 Jun 2015 18:42:00 GMT
Thanks, Supun, great to see your concrete evaluations of this approach.  Is it also possible
to link documents so that frequently changing parts of the data model can be separate docs
and don’t need to be embedded in larger documents?

What other ways can we concretely compare this to the JPA/MySQL approach we use now?

Marlon


From: Supun Nakandala <supun.nakandala@gmail.com<mailto:supun.nakandala@gmail.com>>
Reply-To: "dev@airavata.apache.org<mailto:dev@airavata.apache.org>" <dev@airavata.apache.org<mailto:dev@airavata.apache.org>>
Date: Sunday, June 7, 2015 at 1:49 PM
To: "dev@airavata.apache.org<mailto:dev@airavata.apache.org>" <dev@airavata.apache.org<mailto:dev@airavata.apache.org>>
Subject: Re: Airavata Registry Considerations

Hi All,

I did the initial POC on the subject and would like to summarize the findings here.

I used the same Registry CPI and RegistryImpl class tried to use a mongo db back end  instead
of the JPA one. The architecture of the module is as follows
[Inline image 1]
MongoDB stores its internal data in JSON format and we can simply insert JSON data directly.
Therefore I have used a Thrift to JSON conversion layer and get rid of the additional DB models.
The conversion layer is based on a generic serializer/ deserializer and therefore it is easy
to make data model changes without changing registry. But if we make changes to ID fields
and Indexed -fields those changes has to be updated in the MongoDB indexes. I see getting
rid of the additional DB model classes layer as a -plus point. This also reduces the developer
effort required when incorporating changes.

One major difference between the current registry data modelling and this approach is the
Experiment model. In the MongoDB based model all related data to an experiment is stored in
the same experiment document. A sample JSON which gets stored in the database would be similar
to this https://gist.github.com/scnakandala/19fe3c6edf3be3354439. It is possible to change
sub document contents and retrieve sub documents only. eg. retrieving a specific task. But
in this POC I have not used that function. When updating I have retrieved the entire document
update the required fields in the application logic and update the entire document. The reason
for this is to keep the Dao objects simple as possible and to do the quickly do the implementation.
However it is said the first approach has slight performance advantages compared to the current
approach.

I have push the changes to Airavata git repository under the branch name "mongo-registry".
The implementation is not 100% complete yet but captures most of the idea.

Things I have not investigated yet

1. Read/Write performance
2. Cluster deployment of MongoDB and Consistency of data

Thanks
Supun




On Fri, May 22, 2015 at 10:33 PM, Suresh Marru <smarru@apache.org<mailto:smarru@apache.org>>
wrote:
Hi Supun & Supun :)

This is good discussion. I think we need to balance both aspects here. I am not at all favoring
shoehorning into mongodb and again spend few months addressing the unknowns. On the other
hand, GSoC is the right time to explore alternatives.

My expectation from this document was not so much of criticizing the current JPA based implementation.
Back then the focus was to adopt thrift for the data models (thanks Supun K for the recommendation).
Among other things, thrift helped us to keep the focus on airavata’s core capabilities and
quickly unify all the legacy interfaces. The currently JPA registry was developed from scratch
in a hurry to help with thrift adoption. I think it did well and exceeded initial expectations.

We now slowly circled through all the components and made tremendous progress. We reduced
the internal footprint significantly (rabbitmq in favor of WS Messenger, work queues in place
of custom co-ordination in workflow interpreter and so forth). I think its time to step back
and re-look at the metadata management needs.

How about we not worry on the implementation costs and focus on what criteria we should look
into potential solutions and how to profile them? We should also include the full JPA based
implementation as one of the candidates. As both of you said, its important to identify the
profiling criteria. Chathuri has early work on this, in both the survey paper and performance
measurements, we probably should revisit them and build from there.

Thanks,
Suresh

On May 22, 2015, at 12:28 PM, Supun Nakandala <supun.nakandala@gmail.com<mailto:supun.nakandala@gmail.com>>
wrote:

Hi Supun,

On Fri, May 22, 2015 at 9:42 PM, Supun Kamburugamuve <supun06@gmail.com<mailto:supun06@gmail.com>>
wrote:
Hi Supun,

In normal software developments, it is normal to have these kind of slowness. We cannot foresee
all the things when we develop. The solution is to improve the performance of important operations
rather than re-writing everything from the beginning. For example for this particular select
operations you can directly use SQL rather than going through JPA.

I'm pretty sure you'll encounter more problems, if you implement this in MongoDB than in the
current MySQL. If that happens, do you think abandoning that technology and going for a new
database will be a good solution? Now you have more experience with MySQL than MongoDB as
well.

Rather than going to abandon everything you have because of one problem, trying to fix it
may be better for you in the long run.

Thanks,
Supun..

I completely agree with you. Writing things from scratch will need more development effort
and proper testing. And has the potential of incorporating new unknown issues. It is completely
possible to fix these issues in current registry and I have mentioned that in the doc also.

In addition to that I also checked several other alternatives and found MongoDB interesting.
I am not saying that we should completely rewrite registry using MongoDB. But I think it is
worth exploring it at a POC level.

On Fri, May 22, 2015 at 11:49 AM, Supun Nakandala <supun.nakandala@gmail.com<mailto:supun.nakandala@gmail.com>>
wrote:
Hi Supun,

I haven't done done profiling of registry based operations. Here what I mean by slow performance
is mainly the slowness of the SELECT operations in PHP Reference Gateway. e.g fetching Projects,
fetching experiments. Even a simple query to fetch the 20 most recent experiments is embarrassingly
slow in PGA.

Even though I didn't do a proper profiling of operations I did a query log analysis for a
SELECT experiment query. This was a simple query to fetch 20 most recent experiments. I found
that JPA layer is generating enormous amount of queries for this task rather than one single
query (due to the select N+1 isssue). This issue is same for fetching a single experiment
by specifying the id.

I think it is ok to say that current registry has become bottleneck for most of the PGA specific
operations. But I don't have evidence to show how it has become a bottleneck for the Orchestrator
or GFac specific operations. For that as you have mentioned we need to profile the operations.
But I think the argument is still valid even for GFac and Orchestrator based operations.

I have attached the query log for the above mentioned select operation here with. If you observe
the query log you can see that every associated entity is fetched separately using complex
join operations.



On Fri, May 22, 2015 at 8:05 PM, Supun Kamburugamuve <supun06@gmail.com<mailto:supun06@gmail.com>>
wrote:
Hi Supun,

In your report it says Slow performance. Do you have any data about this slow performance?
For a typical request in what percent the registry slow down the processing compared to overall
time it takes to execute that request?

Do you have a use case where registry is the bottleneck?

Thanks,
Supun..

On Fri, May 22, 2015 at 9:45 AM, Suresh Marru <smarru@apache.org<mailto:smarru@apache.org>>
wrote:
Hi Supun,

This is very good analysis, you have nicely embraced the problem. Before we jump into the
solution, we may want to do small POC’s to validate your claims.

Thank you for getting a headstart, this also cuts into GSoC goals of Douglas’s project.
So lets work on this collaboratively.

Hi Madhu,

Can you please provide guidance on this effort on how to academically approach the data management
challenges of Airavata. The students might appreciate insights on how to profile and benchmark
any possible solutions.

Cheers,
Suresh

On May 22, 2015, at 9:18 AM, Supun Nakandala <supun.nakandala@gmail.com<mailto:supun.nakandala@gmail.com>>
wrote:

Hi Devs,

I have compiled a document based on the analysis I did on current registry architecture/technology
and possible modification and alternatives. You can find the document at https://docs.google.com/document/d/1XWAQLtdtCf9nTigAz6r5JINHR99bP0oeYaTgeEIVr4w/edit#

Thanks
Supun




--
Supun Kamburugamuva
Member, Apache Software Foundation; http://www.apache.org<http://www.apache.org/>
E-mail: supun06@gmail.com<mailto:supun06@gmail.com>;  Mobile: +1 812 369 6762<tel:%2B1%20812%20369%206762>
Blog: http://supunk.blogspot.com<http://supunk.blogspot.com/>




--
Thank you
Supun Nakandala
Dept. Computer Science and Engineering
University of Moratuwa



--
Supun Kamburugamuva
Member, Apache Software Foundation; http://www.apache.org<http://www.apache.org/>
E-mail: supun06@gmail.com<mailto:supun06@gmail.com>;  Mobile: +1 812 369 6762<tel:%2B1%20812%20369%206762>
Blog: http://supunk.blogspot.com<http://supunk.blogspot.com/>




--
Thank you
Supun Nakandala
Dept. Computer Science and Engineering
University of Moratuwa




--
Thank you
Supun Nakandala
Dept. Computer Science and Engineering
University of Moratuwa

Mime
View raw message