airavata-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Amila Jayasekara <>
Subject Re: Work Stealing is not a good solution for Airavata.
Date Wed, 05 Oct 2016 20:43:49 GMT
On Wed, Oct 5, 2016 at 11:14 AM, Shameera Rathnayaka <
> wrote:

> Hi Amila,
> Yeah, I agreed those people who are not much familiar with Airavata
> current internal architecture (as it getting change)  above summary is not
> enough to understand the issue. Let me explain some of your concern with
> more details. Please see my comment inline.
> On Tue, Oct 4, 2016 at 10:51 PM Amila Jayasekara <>
> wrote:
> First of all, what do you mean by "work stealing design pattern"? I have
> not heard such a design pattern and probably what you are trying to explain
> is not "work stealing" (to my understanding). Work stealing is a
> multi-threaded scheduling mechanism mostly use in systems that implement
> lightweight threads. The basic idea is to have a queue per each lightweight
> thread, and when a particular thread is idle, it can get work (steal) from
> a busy lightweight thread. The architecture scenario you are explaining
> does not go into that category of work stealing as per my understanding
> (See [2] for more info). It is a more likely distribution of work.
> Yes, this is not an actually a design pattern, even though I have used it
> in that way. let me take it back and rephrase it, we have used work
> stealing strategy to accomplish and solve distributes system nature and its
> issues. I am well aware what is work stealing is, but Airavata uses that
> concept to distribute work among workers( This is like load balance with
> fixed configurable load per Worker) and also use this to address fault
> tolerance of the components. That means, this work queues are not just a
> tool, it is part of the architecture. In Short, instead of messaging ,
> Airavata have used this AMQP to handle distributes system problems too.

-- Please explain how you used "work stealing" in distributed system. That
would be interesting.

> Regarding 1 => To me, this looks like a limitation in the tools that we
> use. As per what I understood the M, N issue is because of use of AMQP
> prefetch. However, I did not understand, what AMQP prefetch is doing. Just
> because the underlying tools has limitations, I don't think we can
> criticise the design. Always try to keep the conceptual design and tools
> separately.
> Yeah, you are correct, prefetch is something comes with AMQP, and the bad
> thing is Airavata depends on that limit, which decides the maximum load can
> be handled by a worker at a given time. Refer this [1] to get an idea about
> prefetch. If the architecture solely based on AMQP(this is not a tool) to
> distribute works and address FT of Workers then we can't just ignore it.

-- I dont see AMQP in the architecture diagram you attached above and I
dont understand why Airavata has to depend on it. One way to figure this
out is think about the architecture without AMQP and figure out what
actually should happend and look for a way to do that using AMQP.

> Regarding 2 => For cancel experiments keeping priority queue make sense
> but how long a request reside in the worker queue?
> I have no idea what you are trying to explain using zookeeper and rabbitmq
> in this context (also I don't know how those function). If you conceptually
> describe the problem, i may able to give more feedback.
> The worker doesn't ack the message until it finishes the work. The time
> needs for a worker to finish the work is depend on the application run time
> (applications runs on HPC machine). Theoretically, this can be from few sec
> to days or even more.

-- Does this mean that you have a waiting thread or process within Airavata
after submitting the job (for each work) ?

It takes more time for me to digest following right now. I will try to give
more feedback when I properly understand them.


> That means, If a worker can read N(=prefetch count) number of messages
> from a queue without sending acknowledgment then that is the limit one
> worker can handle for given time. But most of this long-running jobs are
> asynchronous. Worker resources are free to handle more works than N. Hence
> Airavata underutilized worker resources. In the case of small jobs (small
> in runtime), this won't be a big problem.
> Apache Zookeeper[2] provide a way to manage distributed system components
> and most of  the latest distributed systems have been used Zookeeper to
> address all the common distributed system problems like HA, FT, Leader
> Election etc ... But in Airavata is trying to replace Rabbitmq with
> Zookeeper to achieve the same outcomes, I haven't seen any framework have
> done it. How Airavata tries to do is using Work Stealing Queues. Anyway,
> Airavata hasn't move zookeeper out of its architecture yet as it uses
> zookeeper to handle cancel requests.
> Regarding 3 => Well.. email monitoring was alway problematic. More
> precisely monitoring is problematic IMO. To handle monitoring I think we
> should be able to get better feedback from job schedulers but as per my
> experience, even those job schedulers are unreliable. Until we have a
> better way to get feedback from job scheduler, monitoring is going to be
> challenging. However, I don't understand why you have "serious scalability"
> issues in GFac because of this.
> Let me explain more about this, I have used some terms comes with latest
> AMQP spec here. Let's say we have two GFac Workers in the system. and both
> submit jobs to Computer Resources and waiting for status update via emails
> (We finally decided to depend on emails for a while come up with more
> robust monitoring solution) when emails come, there is email monitoring
> server which reads this emails and put it to a rabbitmq exchange[3].  Then
> each gfac worker has subscribed to this exchange to get all email updates.
> Because Framework doesn't know which worker handle particular jobs, it
> sends this email content to all workers who subscribe to that exchange.
> There are few issues in this way.
> 1. what if one of the workers goes offline for a moment? It should receive
> the messages from where he left the queue. To do that we need to use
> persistence queue which doesn't remove when consumer disconnect. And this
> queue should receive all email updates messages as well during the consumer
> down time. To create a persistent queue and reconnect again to the same
> queue, the consumer should know about the queue name. Ok, let's say worker
> create this queue  with a name in the very first time it joins to email
> monitoring exchange. Now this problem is solved see the second issue.
> 2. What if one worker node goes down and we start a different node ( runs
> in different Machine/VM) . Now there is no way this new worker knows the
> queue name created by the previous worker unless we configure it which is
> not a very good solution where we have a pool of workers and this pool
> getting changes time to time. Now the real problem is all the jobs handled
> by down node is getting to this new node but there is no way it gets
> previous email monitoring messages. which make these jobs hanging on it
> previous state forever. Even the previously down worker comes up this might
> not get previous jobs instead it retrieves a new set of jobs. This means we
> can't scale Gfac workers independently. Hope this will explain the issue.
> In summary, to me, none of these are concerns for the architecture at the
> moment. Also, you cannot just naively complain an architecture is "not
> good". Architecture has to be compared with another design and evaluate and
> pros/cons for both. I suggest we first try to improve the existing design
> to handle issues you are pointing.
> It would be great to get some solution for above issues. IMHO, we have
> overcomplicated Airavata design with Work Stealing approach, which is not
> suitable for Airavata use cases and requirement.
> Thank you for your feedbacks, hope I answered to all of your concerns,
> [1]
> [2]
> [3]
> Thanks,
> Shameera.
> [2]
> Thanks
> -Amila
> On Tue, Oct 4, 2016 at 1:07 PM, Shameera Rathnayaka <
>> wrote:
> Hi Devs,
> Airavata has adopted to work stealing design pattern lately and use work
> queue approach to distributing works among consumers. There are two work
> queues in current Airavata architecture. One in middle of API Server and
> Orchestrator and the second one in between Orchestrator and Gfac, Following
> is very high-level Airavata architecture.
> [image: Highleve Arachitecture- Airavata.png]
> Here are the issues we have with above architecture.
> 1. Low resource utilization in Workers (Gfac/Orchestrator).
> We have used AMPQ prefetch count to limit the number of requests served by
> a Worker, which is not a good way to load balance in the
> heterogeneous environment where different workers have different level of
> resources. And it is recommended to keep this prefetch count minimum [1]
> and this is valid for work stealing too. If we only have one worker and we
> have M ( > N) number of long running jobs, and our prefetch count is N
> then, only N jobs will in active mode. As we are run this long-running job
> in the async way, we can handle more long running jobs than N. Therefore
> workers resources are underutilized.
> 2. Even though we can easily deal with recovery requirement with work
> stealing, it is not easy to handle cancel experiments. When this cancel
> experiment comes the worker who works on this experiment should act
> immediately. To add this behavior we need to introduce priority queues and
> no need say this will add extra layer of complexity. Currently, we use
> zookeeper to trigger cancel requests ( Another downside, we are using both
> zookeeper and rabbitmq to solve different parts of Distributed systems
> issues. Almost all latest Distributed system frameworks have being used
> zookeeper to manage distributed system problems, we need to strongly
> consider using zookeeper  as a way of managing our components and share the
> load according to the resource available in workers)
> 3. Putting email to a queue is not a good solution with commodity servers
> where system failures are expected. This email queue is critical, if we
> missed one of the statuses of a job then this job can go to the unknown
> state or hang in the old status forever. Due to this, we have serious
> scalability issue with GFac at the moment due to a bottleneck of email
> monitoring.
> I think we need to re-evaluate Airavata architecture and find a good yet
> simple solution based on requirements. The new architecture should handle
> all existing issues and able to extend future requirement.
> [1]
> prefetch-count-value-for-RabbitMQ
> --
> Shameera Rathnayaka
> --
> Shameera Rathnayaka

View raw message