Re: Reusing jobs
When there are non daemon threads, JMX threads being our #1 cause, the jvm will not exit with out help. This is in TaskTracker.java, in 0.16.0, this is line 2088, in the finally clause of Child.main LogManager.shutdown(); System.exit( 0 ); // Force the jvm to exit even if it has threads still running, this prevents memory expensive jvms being left around Devaraj Das wrote: Jason, didn't get that. The jvm should exit naturally even without calling System.exit. Where exactly did you insert the System.exit? Please clarify. Thanks! -Original Message- From: Jason Venner [mailto:[EMAIL PROTECTED] Sent: Friday, April 18, 2008 6:48 PM To: core-user@hadoop.apache.org Subject: Re: Reusing jobs We have terrible issues with threads in the JVM's holding down resources and causing the compute nodes to run out of memory and lock up. We in fact patch the JobTracker to cause the mapper/reduce jvm to System.exit, to ensure that the resources are freed. This is particularly a problem for mapper/reducers that enable jmx or spool off many threads for internal processing. Our solution is to tune the input split size so that the minimum mapper time is > 1 minute Karl Wettin wrote: Ted Dunning skrev: Hadoop has enormous startup costs that are relatively inherent in the current design. Most notably, mappers and reducers are executed in a standalone JVM (ostensibly for safety reasons). Is it possible to hack in support to reuse JVMs? Keep it alive until timed out and have it execute the jobs by opening a socket and say hello? What classes should I start looking in? Could be a fun exercise. karl On 4/17/08 6:00 PM, "Karl Wettin" <[EMAIL PROTECTED]> wrote: Is it possible to execute a job more than once? I use map reduce when adding a new instance to a hierarchial cluster tree. It finds the least distant node and inserts the new instance as a sibling to that node. As far as I know it is in very the nature of this algorithm that one inserts one instance at a time, that this is how the second dimension is created that makes it better than a vector cluster. It would be possible to map all permutations of instances and skip the reduction, but that would result in many more calulations than iteratively training the tree as the latter only require one to test against the instances already inserted to the tree. Iteratively training this tree using Hadoop means executing one job per instance that measure distance to all instances in a file that I also append the new instance to once inserted in the tree. All of above is very inefficient, especially with a young tree that could be trained in nanoseconds locally. So I do that until it takes 20 seconds to insert an instance. But really, this is all Hadoop framework overhead. I'm not quite sure of all it does when I execute a job, but it seems like quite a lot. And all I'm doing is executing a couple of identical jobs over and over again using new data. It would be very nice if I it just took a few milliseconds to do that. karl -- Jason Venner Attributor - Publish with Confidence <http://www.attributor.com/> Attributor is hiring Hadoop Wranglers, contact if interested
RE: Reusing jobs
Jason, didn't get that. The jvm should exit naturally even without calling System.exit. Where exactly did you insert the System.exit? Please clarify. Thanks! > -Original Message- > From: Jason Venner [mailto:[EMAIL PROTECTED] > Sent: Friday, April 18, 2008 6:48 PM > To: core-user@hadoop.apache.org > Subject: Re: Reusing jobs > > We have terrible issues with threads in the JVM's holding > down resources and causing the compute nodes to run out of > memory and lock up. We in fact patch the JobTracker to cause > the mapper/reduce jvm to System.exit, to ensure that the > resources are freed. > > This is particularly a problem for mapper/reducers that > enable jmx or spool off many threads for internal processing. > > Our solution is to tune the input split size so that the > minimum mapper time is > 1 minute > > Karl Wettin wrote: > > Ted Dunning skrev: > >> Hadoop has enormous startup costs that are relatively > inherent in the > >> current design. > >> > >> Most notably, mappers and reducers are executed in a > standalone JVM > >> (ostensibly for safety reasons). > > > > Is it possible to hack in support to reuse JVMs? Keep it > alive until > > timed out and have it execute the jobs by opening a socket and say > > hello? What classes should I start looking in? Could be a > fun exercise. > > > > > > karl > > > > > > > >> > >> > >> > >> On 4/17/08 6:00 PM, "Karl Wettin" <[EMAIL PROTECTED]> wrote: > >> > >>> Is it possible to execute a job more than once? > >>> > >>> I use map reduce when adding a new instance to a > hierarchial cluster > >>> tree. It finds the least distant node and inserts the new > instance > >>> as a sibling to that node. > >>> > >>> As far as I know it is in very the nature of this > algorithm that one > >>> inserts one instance at a time, that this is how the second > >>> dimension is created that makes it better than a vector > cluster. It > >>> would be possible to map all permutations of instances > and skip the > >>> reduction, but that would result in many more calulations than > >>> iteratively training the tree as the latter only require > one to test > >>> against the instances already inserted to the tree. > >>> > >>> Iteratively training this tree using Hadoop means > executing one job > >>> per instance that measure distance to all instances in a > file that I > >>> also append the new instance to once inserted in the tree. > >>> > >>> All of above is very inefficient, especially with a young > tree that > >>> could be trained in nanoseconds locally. So I do that > until it takes > >>> 20 seconds to insert an instance. > >>> > >>> But really, this is all Hadoop framework overhead. I'm not quite > >>> sure of all it does when I execute a job, but it seems > like quite a > >>> lot. And all I'm doing is executing a couple of identical > jobs over > >>> and over again using new data. > >>> > >>> It would be very nice if I it just took a few > milliseconds to do that. > >>> > >>> > >>>karl > >> > > >
Re: Reusing jobs
We have terrible issues with threads in the JVM's holding down resources and causing the compute nodes to run out of memory and lock up. We in fact patch the JobTracker to cause the mapper/reduce jvm to System.exit, to ensure that the resources are freed. This is particularly a problem for mapper/reducers that enable jmx or spool off many threads for internal processing. Our solution is to tune the input split size so that the minimum mapper time is > 1 minute Karl Wettin wrote: Ted Dunning skrev: Hadoop has enormous startup costs that are relatively inherent in the current design. Most notably, mappers and reducers are executed in a standalone JVM (ostensibly for safety reasons). Is it possible to hack in support to reuse JVMs? Keep it alive until timed out and have it execute the jobs by opening a socket and say hello? What classes should I start looking in? Could be a fun exercise. karl On 4/17/08 6:00 PM, "Karl Wettin" <[EMAIL PROTECTED]> wrote: Is it possible to execute a job more than once? I use map reduce when adding a new instance to a hierarchial cluster tree. It finds the least distant node and inserts the new instance as a sibling to that node. As far as I know it is in very the nature of this algorithm that one inserts one instance at a time, that this is how the second dimension is created that makes it better than a vector cluster. It would be possible to map all permutations of instances and skip the reduction, but that would result in many more calulations than iteratively training the tree as the latter only require one to test against the instances already inserted to the tree. Iteratively training this tree using Hadoop means executing one job per instance that measure distance to all instances in a file that I also append the new instance to once inserted in the tree. All of above is very inefficient, especially with a young tree that could be trained in nanoseconds locally. So I do that until it takes 20 seconds to insert an instance. But really, this is all Hadoop framework overhead. I'm not quite sure of all it does when I execute a job, but it seems like quite a lot. And all I'm doing is executing a couple of identical jobs over and over again using new data. It would be very nice if I it just took a few milliseconds to do that. karl
Re: Reusing jobs
Hi -- Not really sure that JVM startup is the main overhead -- you could take a look at the logfiles of the individual TIPs and compare the timestamp of the first log message to the time the jobtracker reports that TIP was started. In my experience, that is well under a second (once the cluster has warmed up), but please do correct me if I'm wrong -- I'd really be interested to know what others observe. BTW, some very rough benchmarks on something similar: http://www.cs.cmu.edu/~spapadim/hadoop/timeline.html The last plot shows executing the job locally (with a chunk size of 128MB) vs a hand-coded C++ program -- both do a simple regex match and then construct a histogram of counts of the matched strings. The overhead is impressively small -- I'm assuming that local execution of a Hadoop job will still fire up a separate JVM for each map chunk (I didn't double-check this). Cheers, Spiros On Thu, Apr 17, 2008 at 10:43 PM, Karl Wettin <[EMAIL PROTECTED]> wrote: > Ted Dunning skrev: > > > Hadoop has enormous startup costs that are relatively inherent in the > > current design. > > > > Most notably, mappers and reducers are executed in a standalone JVM > > (ostensibly for safety reasons). > > > > Is it possible to hack in support to reuse JVMs? Keep it alive until timed > out and have it execute the jobs by opening a socket and say hello? What > classes should I start looking in? Could be a fun exercise. > > > karl > > > > > > > > > > > On 4/17/08 6:00 PM, "Karl Wettin" <[EMAIL PROTECTED]> wrote: > > > > Is it possible to execute a job more than once? > > > > > > I use map reduce when adding a new instance to a hierarchial cluster > > > tree. It finds the least distant node and inserts the new instance as > > > a > > > sibling to that node. > > > > > > As far as I know it is in very the nature of this algorithm that one > > > inserts one instance at a time, that this is how the second dimension > > > is > > > created that makes it better than a vector cluster. It would be > > > possible > > > to map all permutations of instances and skip the reduction, but that > > > would result in many more calulations than iteratively training the > > > tree > > > as the latter only require one to test against the instances already > > > inserted to the tree. > > > > > > Iteratively training this tree using Hadoop means executing one job > > > per > > > instance that measure distance to all instances in a file that I also > > > append the new instance to once inserted in the tree. > > > > > > All of above is very inefficient, especially with a young tree that > > > could be trained in nanoseconds locally. So I do that until it takes > > > 20 > > > seconds to insert an instance. > > > > > > But really, this is all Hadoop framework overhead. I'm not quite sure > > > of > > > all it does when I execute a job, but it seems like quite a lot. And > > > all > > > I'm doing is executing a couple of identical jobs over and over again > > > using new data. > > > > > > It would be very nice if I it just took a few milliseconds to do that. > > > > > > > > > karl > > > > > > > >
Re: Reusing jobs
Ted Dunning skrev: Hadoop has enormous startup costs that are relatively inherent in the current design. Most notably, mappers and reducers are executed in a standalone JVM (ostensibly for safety reasons). Is it possible to hack in support to reuse JVMs? Keep it alive until timed out and have it execute the jobs by opening a socket and say hello? What classes should I start looking in? Could be a fun exercise. karl On 4/17/08 6:00 PM, "Karl Wettin" <[EMAIL PROTECTED]> wrote: Is it possible to execute a job more than once? I use map reduce when adding a new instance to a hierarchial cluster tree. It finds the least distant node and inserts the new instance as a sibling to that node. As far as I know it is in very the nature of this algorithm that one inserts one instance at a time, that this is how the second dimension is created that makes it better than a vector cluster. It would be possible to map all permutations of instances and skip the reduction, but that would result in many more calulations than iteratively training the tree as the latter only require one to test against the instances already inserted to the tree. Iteratively training this tree using Hadoop means executing one job per instance that measure distance to all instances in a file that I also append the new instance to once inserted in the tree. All of above is very inefficient, especially with a young tree that could be trained in nanoseconds locally. So I do that until it takes 20 seconds to insert an instance. But really, this is all Hadoop framework overhead. I'm not quite sure of all it does when I execute a job, but it seems like quite a lot. And all I'm doing is executing a couple of identical jobs over and over again using new data. It would be very nice if I it just took a few milliseconds to do that. karl
Re: Reusing jobs
Hadoop has enormous startup costs that are relatively inherent in the current design. Most notably, mappers and reducers are executed in a standalone JVM (ostensibly for safety reasons). On 4/17/08 6:00 PM, "Karl Wettin" <[EMAIL PROTECTED]> wrote: > Is it possible to execute a job more than once? > > I use map reduce when adding a new instance to a hierarchial cluster > tree. It finds the least distant node and inserts the new instance as a > sibling to that node. > > As far as I know it is in very the nature of this algorithm that one > inserts one instance at a time, that this is how the second dimension is > created that makes it better than a vector cluster. It would be possible > to map all permutations of instances and skip the reduction, but that > would result in many more calulations than iteratively training the tree > as the latter only require one to test against the instances already > inserted to the tree. > > Iteratively training this tree using Hadoop means executing one job per > instance that measure distance to all instances in a file that I also > append the new instance to once inserted in the tree. > > All of above is very inefficient, especially with a young tree that > could be trained in nanoseconds locally. So I do that until it takes 20 > seconds to insert an instance. > > But really, this is all Hadoop framework overhead. I'm not quite sure of > all it does when I execute a job, but it seems like quite a lot. And all > I'm doing is executing a couple of identical jobs over and over again > using new data. > > It would be very nice if I it just took a few milliseconds to do that. > > >karl