In this case the problem is not DUCC, rather it is the high overhead of opening small files and sending them to a remote computer individually. I/O works much more efficiently with larger blocks of data. Many small files can be merged into larger files using zip archives. DUCC sample code shows how to do this for CASes, and very similar code could be used for input documents as well.
Implementing efficient scale out is highly dependent on good treatment of input and output data. Best, Eddie On Sat, Jun 13, 2020 at 6:24 AM Dr. Raja M. Suleman < raja.m.sulai...@gmail.com> wrote: > Hello, > > Thank you very much for your response and even more so for the detailed > explanation. > > So, if I understand it correctly, DUCC is more suited for scenarios where > we have large input documents rather than many small ones? > > Thank you once again. > > On Fri, 12 Jun 2020, 22:18 Eddie Epstein, <eaepst...@gmail.com> wrote: > > > Hi, > > > > In this simple scenario there is a CollectionReader running in a > JobDriver > > process, delivering 100K workitems to multiple remote JobProcesses. The > > processing time is essentially zero. (30 * 60 seconds) / 100,000 > workitems > > = 18 milliseconds per workitem. This time is roughly the expected > overhead > > of a DUCC jobDriver delivering workitems to remote JobProcesses and > > recording the results. DUCC jobs are much more efficient if the overhead > > per workitem is much smaller than the processing time. > > > > Typically DUCC jobs would be processing much larger blocks of content per > > workitem. For example, if a workitem was a document, and the document > > parsed into the small CASes by the CasMultiplier, the throughput would be > > much better. However, with this example, as the number of working > > JobProcess threads is scaled up, the CR (JobDriver) would become a > > bottleneck. That's why a typical DUCC Job will not send the Document > > content as a workitem, but rather send a reference to the workitem > content > > and have the CasMultipliers in the JobProcesses read the content directly > > from the source. > > > > Even though content read by the JobProcesses is much more efficient, as > > scaleout continued to increase for this non-computation scenario the > > bottleneck would eventually move to the underlying filesystem or whatever > > document source and JobProcess output are. The main motivation for DUCC > was > > jobs similar to those in the DUCC examples which use OpenNLP to process > > large documents. That is, jobs where CPU processing is the bottleneck > > rather than I/O. > > > > Hopefully this helps. If not, happy to continue the discussion. > > Eddie > > > > On Fri, Jun 12, 2020 at 1:16 PM Dr. Raja M. Suleman < > > raja.m.sulai...@gmail.com> wrote: > > > > > Hi, > > > Thank you for your reply and I'm sorry I couldn't get back to this > > > earlier. > > > > > > To get a better picture of the processing speed of DUCC, I made a dummy > > > pipeline where the CollectionReader runs a for loop to generate 100k > > > workitems (so no disk reads). each workitem only has a simple string in > > it. > > > These are then passed on to the CasMultiplier where for each workitem > I'm > > > creating a new CAS with DocumentInfo (again only having a simple string > > > value) and pass it as a newcas to the CasConsumer. The CasConsumer > > doesn't > > > do anything except add the Document received in the CAS to the logger. > So > > > basically this pipeline isn't doing anything, no Input reads and the > only > > > output is the information added to the logger. Running this on the > > cluster > > > with 2 slave nodes with 8-CPUs and 32GB RAM each is still taking more > > than > > > 30 minutes. I don't understand how is this possible since there's no > > heavy > > > I/O processing is happening in the code. > > > > > > Any ideas please? > > > > > > Thank you. > > > > > > On 2020/05/18 12:47:41, Eddie Epstein <eaepst...@gmail.com> wrote: > > > > Hi, > > > > > > > > Removing the AE from the pipeline was a good idea to help isolate the > > > > bottleneck. The other two most likely possibilities are the > collection > > > > reader pulling from elastic search or the CAS consumer writing the > > > > processing output. > > > > > > > > DUCC Jobs are a simple way to scale out compute bottlenecks across a > > > > cluster. Scaleout may be of limited or no value for I/O bound jobs. > > > > Please give a more complete picture of the processing scenario on > DUCC. > > > > > > > > Regards, > > > > Eddie > > > > > > > > > > > > On Sat, May 16, 2020 at 1:29 AM Raja Muhammad Suleman < > > > > sulem...@edgehill.ac.uk> wrote: > > > > > > > > > Hi, > > > > > I've been trying to run a very small UIMA DUCC cluster with 2 slave > > > nodes > > > > > having 32GB of RAM each. I wrote a custom Collection Reader to read > > > data > > > > > from an Elasticsearch index and dump it into a new index after > > certain > > > > > analysis engine processing. The Analysis Engine is a simple > sentiment > > > > > analysis code. The performance I'm getting is very slow as it is > only > > > able > > > > > to process ~150 documents/minute. > > > > > To test the performance without the analysis engine, I removed the > AE > > > from > > > > > the pipeline but still I did not get any improvement in the > > processing > > > > > speeds. Can you please guide me as to where I might be going wrong > or > > > what > > > > > I can do to improve the processing speeds? > > > > > > > > > > Thank you. > > > > > ________________________________ > > > > > Edge Hill University<http://ehu.ac.uk/home/emailfooter> > > > > > Teaching Excellence Framework Gold Award< > > > http://ehu.ac.uk/tef/emailfooter> > > > > > ________________________________ > > > > > This message is private and confidential. If you have received this > > > > > message in error, please notify the sender and remove it from your > > > system. > > > > > Any views or opinions presented are solely those of the author and > do > > > not > > > > > necessarily represent those of Edge Hill or associated companies. > > Edge > > > Hill > > > > > University may monitor email traffic data and also the content of > > > email for > > > > > the purposes of security and business communications during staff > > > absence.< > > > > > http://ehu.ac.uk/itspolicies/emailfooter> > > > > > > > > > > > > > > >
