Regards, Andrey
On Wed, Jan 13, 2010 at 10:31 PM, Eric Sammer <[email protected]> wrote: > On 1/13/10 2:09 PM, Andrey Kuzmin wrote: >> On Tue, Jan 12, 2010 at 7:03 PM, Eric Sammer <[email protected]> wrote: >>> On 1/12/10 6:01 AM, Antonio Goncalves wrote: >>>> Thanks Eric and Phil for your inputs. >>>> >>>> We have 80% of our calculation that can be done in one datacenter, but the >>>> rest is heavy calculation. We are using some time consuming algorithm (such >>>> as Monte Carlo for example) that would take too much time in one >>>> datacenter. >>>> For this kind of computation we are thinking of using the second datacenter >>>> based in Germany. We haven't done all the study about the data, but I guess >>>> that for the 80% the data will be local to one datacenter, and for the 20% >>>> it would have to be distributed across datacenter. What we haven't worked >>>> on >>>> yet, is the size of this distributed data. If looks it would not be that >>>> big >>>> (maybe less than a 1Tb but it could grow when doing some calculation based >>>> on archived data). >>> >>> Antonio: >>> >>> The point here is that if you build one logical Hadoop cluster across >>> two data centers, then Hadoop will consider all nodes as candidates for >>> receiving work regardless of which data center the job is started in. >> >> Is Hadoop's job scheduler totally NUMA-unaware? The question holds for >> both cross-data center scenario being discussed and or single data >> center as well: just imagine the usual within-rack or between-racks >> scheduling decision. > > Andrey: > > Take a look at how replicas are assigned to data nodes[1] to see how > blocks are distributed. During M/R the job tracker will assign a map > task to a tracker where the input split is "as close to the data as > possible." Close is either data local (where the task tracker is running > on the same machine as the data), rack local (task tracker on a machine > in the same rack as the data), or not local at all. What I was saying is > that the third case is always possible (and undesirable). Right, but this also means that, with NUMA-aware scheduler, a job distributed between data centers will run as good as it can be achieve wrt locality. Hence, theoretically there could be computation models where cross-data center jobs are feasible (disregarding singe name-node issue) rather than "undesirable". Regards, Andrey > > [1] > http://hadoop.apache.org/common/docs/current/hdfs_design.html#Replica+Placement%3A+The+First+Baby+Steps > > -- > Eric Sammer > [email protected] > http://esammer.blogspot.com >
