airavata-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Shenoy, Gourav Ganesh" <goshe...@indiana.edu>
Subject Re: [#Spring17-Airavata-Courses] : Distributed Workload Management for Airavata
Date Thu, 09 Feb 2017 02:59:49 GMT
Hi All,

As I mentioned before, here is the design we have kind of reached a consensus on (please do
provide comments/suggestions). This idea has been motivated from an understanding of the Aurora/Mesos
architecture, and how they function.

[cid:image001.png@01D28256.A72BF580]

This design has the following benefits:

-          Loosely coupled, independent micro-services.

-          Inherently scalable in nature.

-          Highly available, and consistent architecture.

-          Supports incremental upgrade, without the risk of breaking any existing implementation
while doing so.

-          Ability to add/remove tasks in a DAG, and also add new task implementations (abstraction).

-          Custom scheduler provides us greater flexibility (see below).

We have the orchestrator (will eventually be HA using zookeeper), which will centrally maintain
the state of an experiment – in short the status of the tasks it composes. Based on the
type of job request, it will fetch the task execution DAG – this DAG will be made pre-available
to the orchestrator via a graph database (debatable), and this DAG is nothing but a definition
of sequence of tasks needed for that experiment (not the implementation of tasks).

There is a scheduler which will receive a task execution request from the orchestrator, and
decide which worker will be executing it. each worker here will be analogous to the current
Airavata GFAC module which executes the task. We can think of the worker to be a collection
of implementations of different tasks. Eg: W1, W2, W3 in figure above will have code to execute
tasks A, B, C, D.

There are 2 concerns which arise here:

-          How does the scheduler know/decide which worker to pass on the task execution to?

-          How do we upgrade a worker, say with a new task ‘E’ implementation, in such
a manner that if something goes wrong with code for ‘E’, the entire worker node should
not fail? In short, avoid regression testing the entire worker module.

To address the first problem, I suggest we use a paradigm similar to how Aurora agents (workers)
report available capabilities to the Aurora master (scheduler). In Aurora, the slave nodes
constantly report back to the master how much processing power they have; and accordingly,
the master decides which slave to pass a new job request to. In our case, we can have the
workers advertise to the scheduler which tasks they are capable of executing and the scheduler
acts accordingly.

To address the second concern, I suggest we have the task implementations bundled in separate
JARs, so that if there is a problem with one task the others don’t get affected and can
be “repaired” without impacting other existing tasks impls. There might be better ways
to do this, but this is what I could think of right now.

As mentioned before, adding a new task implementation – which will need upgrades to all
workers will be easy and hassle-free as each worker will report back to the scheduler their
capability to handle that new task, as and when upgrade finishes (incremental upgrade). Having
a custom scheduler also provides us other benefits such as:

-          Handling corner cases – eg: task execution on one worker fails (for some unforeseen
reason), then the scheduler can retry it on a different worker.

-          Prioritize experiments – scheduler higher priority experiments before normal
priority ones (I just made this one up).

We have decided to go ahead and start building a prototype of this design starting tomorrow,
unless there are any concerns/issues. Please do let me know your views on this approach, as
every concern helps us better our design.

Thanks and Regards,
Gourav Shenoy


From: "Shenoy, Gourav Ganesh" <goshenoy@indiana.edu>
Reply-To: "dev@airavata.apache.org" <dev@airavata.apache.org>
Date: Wednesday, February 8, 2017 at 7:06 PM
To: "dev@airavata.apache.org" <dev@airavata.apache.org>
Subject: Re: [#Spring17-Airavata-Courses] : Distributed Workload Management for Airavata

Hi Amruta,

Thanks for providing your inputs, and yes in fact we had started out our design discussions
with a decentralized framework in mind. But then we considered the problem of making each
micro-service independent of each other and more importantly not making them aware of what
the DAG is. For this reason, we decided to push and maintain the DAG at a centralized &
highly available place (the orchestrator), giving us more control and flexibility in adding/removing
tasks from the DAG. This also provides us with the ability to scale each service when needed
and also perform incremental upgrades via devops.

Do let me know if I make sense, or if there is something I am missing. I would also like to
add that we have today nearly come to a consensus on a “fairly good” design – which
I will be detailing in another email shortly.

Thanks and Regards,
Gourav Shenoy

From: "Kamat, Amruta Ravalnath" <arkamat@indiana.edu>
Reply-To: "dev@airavata.apache.org" <dev@airavata.apache.org>
Date: Wednesday, February 8, 2017 at 2:59 AM
To: "dev@airavata.apache.org" <dev@airavata.apache.org>
Subject: Re: [#Spring17-Airavata-Courses] : Distributed Workload Management for Airavata


Hello Gourav,



I agree with your solution, but I just came across a decentralized architecture which might
serve our purpose and might provide a looser coupling.



Having a common workflow would mean a centralized orchestrator i.e. a process which coordinates
with multiple services to complete a larger workflow. The services have no knowledge of the
workflow or their specific involvement in it. The orchestrator takes care of the complexities.
However, The challenge with an orchestrator is that business logic will build up in a central
place.
If there is a central shared instance of the orchestrator for all requests, then the orchestrator
is a single point of failure. If it goes down, all processing stops.

With decentralized interactions, each service takes full responsibility for its role in the
greater workflow. It will listen for events from other services, complete it's work as soon
as possible, retry if a failure occurs and send out events upon completion. Here, communications
tend to be asynchronous and business logic stays within the related services.
Instead of having a central orchestrator that controls the logic of what steps happen when,
that logic is built into each service ahead of time. The services know what to react to and
how, ahead of time. Multiple services can consume the same events, do some processing, and
then produce their own events back into the event stream, all at the same time. The event
stream does not have any logic and is intended to be a dumb pipe.


​Decentralized interactions meet our requirements better: loose coupling, high cohesion
and each service responsible for it's own bounded context.

Thanks
Amruta Kamat
________________________________
From: Shenoy, Gourav Ganesh <goshenoy@indiana.edu>
Sent: Tuesday, February 7, 2017 11:49 PM
To: dev@airavata.apache.org
Subject: Re: [#Spring17-Airavata-Courses] : Distributed Workload Management for Airavata

Supun,

Thank you for this excellent explanation. I see that the architecture you mentioned covers
most of the concerns we discussed in this thread and in class. I just had one clarifying question
though – what does “worker” signify here? Is it a generic task execution framework which
runs the DAG? Or is it a like a platform where the DAG runs (and how?).

Apart from that, I am looking at Storm’s architecture to see if we can get some clues as
they are tackling a similar problem. I shall update once I get some concrete answer.

Thanks and Regards,
Gourav Shenoy

From: Supun Nakandala <supun.nakandala@gmail.com>
Reply-To: "dev@airavata.apache.org" <dev@airavata.apache.org>
Date: Tuesday, February 7, 2017 at 5:47 PM
To: dev <dev@airavata.apache.org>
Subject: Re: [#Spring17-Airavata-Courses] : Distributed Workload Management for Airavata

Hi Gourav,I agree with your idea of using one “workflow micro-service” which would basically
be the mediator/orchestrator for deciding which micro-service should be executed next. But
I think these components do not necessarily have to be micro-services but rather conforms
to the master-worker paradigm in some sense. But the trick here is how can we implement a
scalable, fault tolerant system to do distributed workload management and from CAP theorem
what is the property that we are going to compromise.

I think you are heading in the right direction. But I would like to add more details to your
solution. Please note that I haven't evaluated these ideas 100%. Perhaps we can talk more
about this in the next class.

As you have done, I think we should centralize the state information into one component (orchestrator
in our case). From my experience, it is very hard to achieve consistency in a distributed
state setting in the events of failure.

Second, to maintain generalizability in Airavata I think we should treat each application/use-cases
as a DAG of execution. For example, HPC job and a cloud job will have two different DAGs which
consists of tasks (data staging, job submission, out staging etc). These tasks should be short
tasks and should roughly have the same execution time. And having idempotent tasks is preferable.

Orchestrator is responsible for executing the DAG and assign tasks to the workers(how? will
follow) based on the control dependencies in the DAG tasks. In addition to the dependencies
generated from tasks I see, there can be other dependencies to things like monitoring and
scheduling which the orchestrator has to make into account when executing the DAG.

The next question is how we distribute jobs from Orchestrator to workers. I think here it
is ok to compromise availability in favor of consistency. I suggest that we use the request/response
messaging pattern which uses a persistent message broker (critical service). In this architecture,
we can safely allow orchestrator or workers to fail without losing consistency (because of
the persistent queue). But if the orchestrator fails then the availability will go down. One
way to overcome this would be to come up with an orchestrator quorum.Attached figure summarizes
my idea.

I think we can also evaluate this solution with the concerns that Shameera pointed out such
as can we enable cancel?. Once again it's just my idea and is open for argument and debate.



[ine image 2]

Thanks
-Supun



On Tue, Feb 7, 2017 at 10:54 AM, Shenoy, Gourav Ganesh <goshenoy@indiana.edu<mailto:goshenoy@indiana.edu>>
wrote:
Hi Supun,

I agree, but may be for the example I mentioned, multiple micro-services might not sound necessary.
I was trying to generalize towards a scenario where we have multiple independent micro-services
(not necessarily for task execution). Again, I am not certain if this is the right architecture
but yours (and other’s) inputs, will definitely help us narrow down on the different scenarios
we need to exactly focus on. Do let me know if I make sense.

Thanks and Regards,
Gourav Shenoy

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: Monday, February 6, 2017 at 12:15 PM
To: dev <dev@airavata.apache.org<mailto:dev@airavata.apache.org>>

Subject: Re: [#Spring17-Airavata-Courses] : Distributed Workload Management for Airavata

Hi Gourav,

It is my belief that we don't need a separate microservice to each task. I favor a single
micro service which can execute all tasks (or in other words a generic task execution micro
service). Of course, we can have many of them when we want to scale. WDYT?

On Sun, Feb 5, 2017 at 3:07 PM, Shenoy, Gourav Ganesh <goshenoy@indiana.edu<mailto:goshenoy@indiana.edu>>
wrote:
Hi dev,

We were brainstorming some potential designs that might help us with this problem. One possible
option would be to have a “workflow micro-service” which would basically be the mediator/orchestrator
for deciding which micro-service should be executed next – based on the type of the job.
The motive is to make micro-services independent of the workflow; i.e. a micro-service implementation
should be not be aware of which micro-service will be executed next and we should have a central
control of deciding this pattern.
Eg: For job type X, the pattern could be A -> B -> C -> D. Whereas for job type Y,
the pattern could be A -> C -> D; and so on.

An initial design with this idea looks like follows:
[cid:image003.png@01D28256.A72BF580]


We would have a common messaging framework (implementation has not been decided yet). The
database associated with the workflow micro-service could be a graph database (maybe?) –
again the implementation/technology has not been decided yet.

This is just a proposed design, and I would love to hear your thoughts on this and any suggestions/comments
if any. If there is anything that we are missing or should consider, please do let us know.

Thanks and Regards,
Gourav Shenoy

From: "Christie, Marcus Aaron" <machrist@iu.edu<mailto:machrist@iu.edu>>
Reply-To: "dev@airavata.apache.org<mailto:dev@airavata.apache.org>" <dev@airavata.apache.org<mailto:dev@airavata.apache.org>>
Date: Friday, February 3, 2017 at 9:21 AM

To: "dev@airavata.apache.org<mailto:dev@airavata.apache.org>" <dev@airavata.apache.org<mailto:dev@airavata.apache.org>>
Subject: Re: [#Spring17-Airavata-Courses] : Distributed Workload Management for Airavata

Vidya,

I’m not sure how relevant it is, but it occurs to me that a microservice that executes jobs
on a cloud requires very little in terms of resources to submit and monitor that job on the
cloud. It doesn’t really matter if the job is a “big” or a “small” job.  So I’m
not sure what heuristic makes sense regarding distributing work to these job execution microservices.
 Maybe a simple round robin approach would be sufficient.

I think a job scheduling algorithm does make sense, however, for a higher level component,
some sort of metascheduler that understands what resources are available on the cloud resources
on which the jobs will be running.  The metascheduler could create work for the job exection
microservices to run on particular cloud resources in a way that optimizes for some metric
(e.g., throughput).

Thanks,

Marcus

On Feb 3, 2017, at 3:19 AM, Vidya Sagar Kalvakunta <vkalvaku@umail.iu.edu<mailto:vkalvaku@umail.iu.edu>>
wrote:

Ajinkya,

My scenario is for workload distribution among multiple instances of the same microservice.

If a message broker needs to distribute the available jobs among multiple workers, the common
approach would be to use round robin or a similar algorithm. This approach works best when
all the workers are similar and the jobs are equal.

So I think that a genetic or heuristic job scheduling algorithm, which is also aware of each
of the worker's current state (CPU, RAM, No of Jobs processing) can more efficiently distribute
the jobs. The workers can periodically ping the message broker with their current state info.

The other advantage of using a customized algorithm is that it can be tweaked to use embedded
routing, priority or other information in the job metadata to resolve all of the concerns
raised by Amrutha viz message grouping, ordering, repeated messages, etc.

We can even ensure data privacy, i.e if the workers are spread across multiple compute clusters
say AWS and IU Big Red and we want to restrict certain sensitive jobs to be run only on Big
Red.

Some distributed job scheduling algorithms for cloud computing.

  *   http://www.ijimai.org/journal/sites/default/files/files/2013/03/ijimai20132_18_pdf_62825.pdf
  *   https://arxiv.org/pdf/1404.5528.pdf


Regards
Vidya Sagar

On Fri, Feb 3, 2017 at 1:38 AM, Kamat, Amruta Ravalnath <arkamat@indiana.edu<mailto:arkamat@indiana.edu>>
wrote:
Hello all,

Adding more information to the message based approach. Messaging is a key strategy employed
in many distributed environments. Message queuing is ideally suited to performing asynchronous
operations. A sender can post a message to a queue, but it does not have to wait while the
message is retrieved and processed. A sender and receiver do not even have to be running concurrently.

With message queuing there can be 2 possible scenarios:

  1.  ​Sending and receiving messages using a single message queue.
  2.  ​Sharing a message queue between many senders and receivers
​When a message is retrieved, it is removed from the queue. A message queue may also support
message peeking. This mechanism can be useful if several receivers are retrieving messages
from the same queue, but each receiver only wishes to handle specific messages. The receiver
can examine the message it has peeked, and decide whether to retrieve the message (which removes
it from the queue) or leave it on the queue for another receiver to handle.

A few basic message queuing patterns are:

  1.  One-way messaging: The sender simply posts a message to the queue in the expectation
that a receiver will retrieve it and process it at some point.
  2.  Request/response messaging: In this pattern a sender posts a message to a queue and
expects a response from the receiver. The sender can resend if the message is not delivered.
This pattern typically requires some form of correlation to enable the sender to determine
which response message corresponds to which request sent to the receiver.
  3.  Broadcast messaging: In this pattern a sender posts a message to a queue, and multiple
receivers can read a copy of the message. This pattern depends on the message queue being
able to disseminate the same message to multiple receivers. There is a queue to which the
senders can post messages that include metadata in the form of attributes. Each receiver can
create a subscription to the queue, specifying a filter that examines the values of message
attributes. Any messages posted to the queue with attribute values that match the filter are
automatically forwarded to that subscription.
A solution based on asynchronous messaging might need to address a number of concerns:

Message ordering, Message grouping: Process messages either in the order they are posted or
in a specific order based on priority. Also, there may be occasions when it is difficult to
eliminate dependencies, and it may be necessary to group messages together so that they are
all handled by the same receiver.
Idempotency: Ideally the message processing logic in a receiver should be idempotent so that,
if the work performed is repeated, this repetition does not change the state of the system.
Repeated messages: Some message queuing systems implement duplicate message detection and
removal based on message IDs
Poison messages: A poison message is a message that cannot be handled, often because it is
malformed or contains unexpected information.
Message expiration: A message might have a limited lifetime, and if it is not processed within
this period it might no longer be relevant and should be discarded.
Message scheduling: A message might be temporarily embargoed and should not be processed until
a specific date and time. The message should not be available to a receiver until this time.


Thanks
Amruta Kamat
________________________________
From: Shenoy, Gourav Ganesh <goshenoy@indiana.edu<mailto:goshenoy@indiana.edu>>
Sent: Thursday, February 2, 2017 7:57 PM
To: dev@airavata.apache.org<mailto:dev@airavata.apache.org>

Subject: Re: [#Spring17-Airavata-Courses] : Distributed Workload Management for Airavata

Hello all,

Amila, Sagar, thank you for the response and raising those concerns; and apologies because
my email resonated the topic of workload management in terms of how micro-services communicate.
As Ajinkya rightly mentioned, there exists some sort of correlation between micro-services
communication and it’s impact on how that micro-service performs the work under those circumstances.
The goal is to make sure we have maximum independence between micro-services, and investigate
the workflow pattern in which these micro-services will operate such that we can find the
right balance between availability & consistency. Again, from our preliminary analysis
we can assert that these solutions may not be generic and the specific use-case will have
a big decisive role.

For starters, we are focusing on the following example – and I think this will clarify the
doubts on what we are exactly trying to investigate about.

Our test example
Say we have the following 4 micro-services, which each perform a specific task as mentioned
in the box.

<image001.png>


A state-full pattern to distribute work
<image002.png>

Here each communication between micro-services could be via RPC or Messaging (eg: RabbitMQ).
Obvious disadvantage is that if any micro-service is down, then the system availability is
at stake. In this test example, we can see that Microservice-A coordinates the work and maintains
the state information.

A state-less pattern to distribute work

<image003.png>

Another purely asynchronous approach would be to associate message-queues with each micro-service,
where each micro-service performs it’s task, submits a request (message on bus) to the next
micro-service, and continues to process more requests. This ensures more availability, and
perhaps we might need to handle corner cases for failures such as message broker down, or
message loss, etc.

As mentioned, these are just a few proposals that we are planning to investigate via a prototype
project. Inject corner cases/failures and try and find ways to handle these cases. I would
love to hear more thoughts/questions/suggestions.

Thanks and Regards,
Gourav Shenoy

From: Ajinkya Dhamnaskar <adhamnas@umail.iu.edu<mailto:adhamnas@umail.iu.edu>>
Reply-To: "dev@airavata.apache.org<mailto:dev@airavata.apache.org>" <dev@airavata.apache.org<mailto:dev@airavata.apache.org>>
Date: Thursday, February 2, 2017 at 2:22 AM
To: "dev@airavata.apache.org<mailto:dev@airavata.apache.org>" <dev@airavata.apache.org<mailto:dev@airavata.apache.org>>
Subject: Re: [#Spring17-Airavata-Courses] : Distributed Workload Management for Airavata

Hello all,

Just a heads up. Here the name Distributed workload management does not necessarily mean having
different instances of a microservice and then distributing work among these instances.

Apparently, the problem is how to make each microservice work independently with concrete
distributed communication infrastructure. So, think of it as a workflow where each microservice
does its part of work and communicates (how? yet to be decided) output. The next underlying
microservice identifies and picks up that output and takes it further towards the final outcome,
having said that, the crux here is, none of the miscoservices need to worry about other miscoservices
in a pipeline.

Vidya Sagar,
I completely second your opinion of having stateless miscoservices, in fact that is the key.
With stateless miscroservices it is difficult to guarantee consistency in a system but it
solves the availability problem to some extent. I would be interested to understand what do
you mean by "an intelligent job scheduling algorithm, which receives real-time updates from
the microservices with their current state information".

On Wed, Feb 1, 2017 at 11:48 PM, Vidya Sagar Kalvakunta <vkalvaku@umail.iu.edu<mailto:vkalvaku@umail.iu.edu>>
wrote:

On Wed, Feb 1, 2017 at 2:37 PM, Amila Jayasekara <thejaka.amila@gmail.com<mailto:thejaka.amila@gmail.com>>
wrote:
Hi Gourav,

Sorry, I did not understand your question. Specifically I am having trouble relating "work
load management" to options you suggest (RPC, message based etc.).
So what exactly you mean by "workload management" ?
What is work in this context ?

Also, I did not understand what you meant by "the most efficient way". Efficient interms of
what ? Are you looking at speed ?

As per your suggestions, it seems you are trying to find a way to communicate between micro
services. RPC might be troublesome if you need to communicate with processes separated from
a firewall.

Thanks
-Thejaka


On Wed, Feb 1, 2017 at 12:52 PM, Shenoy, Gourav Ganesh <goshenoy@indiana.edu<mailto:goshenoy@indiana.edu>>
wrote:
Hello dev, arch,

As part of this Spring’17 Advanced Science Gateway Architecture course, we are working on
trying to debate and find possible solutions to the issue of managing distributed workloads
in Apache Airavata. This leads to the discussion of finding the most efficient way that different
Airavata micro-services should communicate and distribute work, in such a way that:

1.       We maintain the ability to scale these micro-services whenever needed (autoscale
perhaps?).

2.       Achieve fault tolerance.

3.       We can deploy these micro-services independently, or better in a containerized manner
– keeping in mind the ability to use devops for deployment.

As of now the options we are exploring are:

1.       RPC based communication

2.       Message based – either master-worker, or work-queue, etc

3.       A combination of both these approaches

I am more inclined towards exploring the message based approach, but again there arises the
possibility of handling limitations/corner cases of message broker such as downtimes (may
be more). In my opinion, having asynchronous communication will help us achieve most of the
above-mentioned points. Another debatable issue is making the micro-services implementation
stateless, such that we do not have to pass the state information between micro-services.

I would love to hear any thoughts/suggestions/comments on this topic and open up a discussion
via this mail thread. If there is anything that I have missed which is relevant to this issue,
please let me know.

Thanks and Regards,
Gourav Shenoy


Hi Gourav,

Correct me if I'm wrong, but I think this is a case of the job shop scheduling problem, as
we may have 'n' jobs of varying processing times and memory requirements, and we have 'm'
microservices with possibly different computing and memory capacities, and we are trying to
minimize the makespan<https://en.wikipedia.org/wiki/Makespan>.

For this use-case, I'm in favor a highly available and consistent message broker with an intelligent
job scheduling algorithm, which receives real-time updates from the microservices with their
current state information.

As for the state vs stateless implementation, I think that question depends on the functionality
of a particular microservice. In a broad sense, the stateless implementation should be preferred
as it will scale better horizontally.


Regards,
Vidya Sagar


--
Vidya Sagar Kalvakunta | Graduate MS CS Student | IU School of Informatics and Computing |
Indiana University Bloomington | (812) 691-5002<tel:8126915002> | vkalvaku@iu.edu<mailto:vkalvaku@iu.edu>



--
Thanks and regards,

Ajinkya Dhamnaskar
Student ID : 0003469679
Masters (CS)
+1 (812) 369- 5416<tel:(812)%20369-5416>



--
Vidya Sagar Kalvakunta | Graduate MS CS Student | IU School of Informatics and Computing |
Indiana University Bloomington | (812) 691-5002<tel:8126915002> | vkalvaku@iu.edu<mailto:vkalvaku@iu.edu>




--
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