There are 2 modes of doing iterations. 1) Client controlled iterations (aka Spark) - where the client runs a DAG with 1 or more iterations and upon their completion, the client determines the termination condition. If the condition is not met then the client submits more DAGs until it is so. With Tez session and shared object model to share data between tasks, this model can be efficiently supported. The final write of each set of iterations has to be in some distributed store, e.g. HDFS. We have seen this model work with comparable performance to Spark (when doing tests on an experimental prototype of Spark with Tez). 1-1 edges can be used to make per vertex iterations really fast.
2) Job controlled iterations (aka Flink) - where the job itself determines the termination state and adds more iterations as needed. This is currently not support in Tez (addition of vertices or early DAG exit) but there are jiras open for those items. Bikas -----Original Message----- From: Johannes Zillmann [mailto:[email protected]] Sent: Wednesday, March 25, 2015 10:24 AM To: [email protected] Subject: Re: Which computation model does Tez supports Hey Hitesh, thanks for you thoughts! In one chooses the multi-vertex approach, i guess there is no simple thing one could do to achieve n-iterations where n is flexible based on the output of the n-1 iteration. So you can't do - do max 20 iterationos - stop in case certain conditions are met !? Johannes > On 25 Mar 2015, at 18:05, Hitesh Shah <[email protected]> wrote: > > Hi Johannes, > > You would likely not avoid it if you went with the approach of multiple DAGs. > For most iterative programs, you do need to checkpoint at some point. The > checkpoint would likely need to be reliable to reduce the amount of > re-computation needed if the check pointed data is lost. An option would be > to use something like the HDFS in-memory storage tier ( which lazily persists > to disk ) to reduce the perf overhead. Also, in terms of loop unrolling, a > single DAG could be pre-constructed to run multiple iterations using multiple > vertices and then use the final vertex of the DAG as a checkpointing > mechanism after N iterations/vertices. > > Also, depending on the amount of data being written out, the overhead of > writing to HDFS may not be too high. Furthermore, with Tez sessions, there is > no real overhead of launching a new DAG ( if some containers are retained ) > as compared to trying to do the same with multiple MR jobs. > > - Hitesh > > > On Mar 25, 2015, at 2:02 AM, Johannes Zillmann <[email protected]> > wrote: > >> Hey Gopal, >> >>> On 25 Mar 2015, at 05:26, Gopal Vijayaraghavan <[email protected]> wrote: >>> >>> Hi, >>> >>> Iterative algorithms are expressed as DAGs in a loop. >>> >>> The acyclic nature of DAGs, whether in Tez or Spark (since you >>> mention the >>> paper) make that the natural way to implement that - repeated >>> application of the same operation over the same data, with a >>> decision condition determining whether to stay in the loop or not. >> >> Can you point to a piece of code which implements this approach ? >> If you each look operation is a single DAG, how would that avoid hdfs >> barrier ? >> >> Johannes >> >>> >>> You might want to look at last year¹s Hadoop Summit presentations >>> for a direct example of Iterative algorithms with Tez. >>> >>> http://www.slideshare.net/Hadoop_Summit/pig-on-tez-low-latency-etl-w >>> ith-big >>> -data/25 >>> >>> >>> Logistic regression needs you to use a library which implements that >>> specific algorithm [1]. >>> >>> On that note, something which needs incremental iteration can >>> probably be even more efficient in Tez than these approaches if you >>> unroll the iteration as 1-1 edges all of the final tasks ending up >>> generating outputs. >>> >>> Cheers, >>> Gopal >>> [1] - https://github.com/myui/hivemall#regression >>> >>> >>> On 3/24/15, 8:43 PM, "Chang Chen" <[email protected]> wrote: >>> >>>> Hi >>>> >>>> from the PhD Disseration >>>> <http://www.eecs.berkeley.edu/Pubs/TechRpts/2014/EECS-2014-12.pdf> >>>> of Matei Zaharia, there are four computation models in the large >>>> scale clusters: >>>> >>>> >>>> 1. *Iterative algorithm*, such as graph processing and machine >>>> leaning algorithm 2. *Relational query* 3. *MapReduce*, a general >>>> parallel computation model 4. *Stream processing*, >>>> >>>> Obviously, Tez supports #2 and #3, but for #1 and #4, I don't see >>>> any examples. >>>> >>>> As for streaming, I guess if we implement appropriate input, there >>>> is no reason that tez can't support in theory. >>>> >>>> But for Machine Leaning, how do we use vertex and edge to express >>>> *Logistic Regression*? >>>> >>>> Thanks >>>> Chang >>> >>> >> >
