Automated load-balancing of partitions in arbitrarily imbalanced distributed mapreduce computations
A distributed computing system executes a MapReduce job on streamed data that includes an arbitrary amount of imbalance with respect to the frequency distribution of the data keys in the dataset. A map task module maps the dataset to a coarse partitioning, and generates a list of the top K keys with the highest frequency among the dataset. A sort task module employs a plurality of sorters to read the coarse partitioning and sort the data into buckets by data key. The values for the top K most frequent keys are separated into single-key buckets. The other less frequently occurring keys are assigned to buckets that each have multiple keys assigned to it. Then, more than one worker is assigned to each single-key bucket. The output of the multiple workers assigned to each respective single-key bucket is stitched together.
1. A computer-implemented method of load-balancing in an arbitrarily imbalanced MapReduce job in a distributed computing system, the method comprising:
identifying K data keys with the highest frequency among received data, the received data comprising pairings of data keys and data values to be processed in the MapReduce job, wherein K is determined according to a data key frequency distribution of the received data and a threshold level of acceptable imbalance in reduce phase worker loads across reduce phase workers, and wherein K is a number selected to keep a maximum imbalance ratio under a threshold;
assigning data for each of the K data keys to a single-key bucket and other data keys to multiple-key buckets;
assigning one respective reduce phase worker to process data values corresponding to data keys of each multiple-key bucket, each multiple-key bucket comprising queued data items having several different keys;
assigning multiple reduce phase workers to process data values corresponding to the data key of each single-key bucket; and
stitching together output of the assigned multiple reduce phase workers on each respective single-key bucket.
2. The method of claim 1 , wherein the received data is a data stream.
3. The method of claim 1 , wherein the identified K keys with the highest frequency among the received data fluctuates as additional data are received.
4. The method of claim 1 , wherein K is a number selected so a frequency of a data key assigned to a single-key bucket with a lowest frequency among data keys assigned to single-key buckets exceeds a threshold.
5. The method of claim 1 , wherein a number of multiple reduce phase workers to assign to a single-key bucket is determined based on a respective frequency of the data key assigned to the single-key bucket and a threshold level of acceptable imbalance in reduce phase worker loads across reduce phase workers.
6. The method of claim 1 , further comprising reporting a frequency distribution of the K highest frequency data keys.
7. The method of claim 1 , further comprising combining output across all reduce phase workers to obtain a result of a MapReduce computation.
8. The method of claim 1 , wherein assigning each of the K data keys to a single-key bucket and other data keys to multiple-key buckets comprises a single sort step.
9. The method of claim 1 , wherein the single-key bucket for a first data key of the identified K data keys is arbitrarily large compared to other buckets.
10. A nontransitory computer readable storage medium including computer program instructions that, when executed, cause a computer processor to perform operations comprising:
identifying K data keys with the highest frequency among received data, the received data comprising pairings of data keys and data values to be processed in the MapReduce job, wherein K is determined according to a data key frequency distribution of the received data and a threshold level of acceptable imbalance in reduce phase worker loads across reduce phase workers, and wherein K is a number selected to keep a maximum imbalance ratio under a threshold;
assigning data for each of the K data keys to a single-key bucket and other data keys to multiple-key buckets;
assigning one respective reduce phase worker to process data values corresponding to data keys of each multiple-key bucket, each multiple-key bucket comprising queued data items having several different keys;
assigning multiple reduce phase workers to process data values corresponding to the data key of each single-key bucket; and
stitching together output of the assigned multiple reduce phase workers on each respective single-key bucket.
11. The medium of claim 10 , wherein the received data is a data stream.
12. The medium of claim 10 , wherein the identified K keys with the highest frequency among the received data fluctuates as additional data are received.
13. The medium of claim 10 , wherein a number of multiple reduce phase workers to assign to a single-key bucket is determined based on a respective frequency of the data key assigned to the single-key bucket and a threshold level of acceptable imbalance in reduce phase worker loads across reduce phase workers.
14. The medium of claim 10 , wherein the operations further comprise reporting a frequency distribution of the K highest frequency data keys.
15. The medium of claim 10 , wherein the operations further comprise combining output across all reduce phase workers to obtain a result of a MapReduce computation.
16. The medium of claim 10 , wherein assigning each of the K data keys to a single-key bucket and other data keys to multiple-key buckets comprises a single sort step.
17. The medium of claim 10 , wherein the single-key bucket for a first data key of the identified K data keys is arbitrarily large compared to other buckets.
18. The medium of claim 10 , wherein K is a number selected so a frequency of a data key assigned to a single-key bucket with a lowest frequency among data keys assigned to single-key buckets exceeds a threshold.
19. A system comprising:
a computer processor; and
a computer readable storage medium storing processor-executable computer program instructions, the computer program instructions comprising instructions for:
identifying K data keys with the highest frequency among received data, the received data comprising pairings of data keys and data values to be processed in the MapReduce job, wherein K is determined according to a data key frequency distribution of the received data and a threshold level of acceptable imbalance in reduce phase worker loads across reduce phase workers, and wherein K is a number selected to keep a maximum imbalance ratio under a threshold;
assigning data for each of the K data keys to a single-key bucket and other data keys to multiple-key buckets;
assigning one respective reduce phase worker to process data values corresponding to data keys of each multiple-key bucket, each multiple-key bucket comprising queued data items having several different keys;
assigning multiple reduce phase workers to process data values corresponding to the data key of each single-key bucket; and
stitching together output of the assigned multiple reduce phase workers on each respective single-key bucket.
20. The system of claim 19 , wherein a number of multiple reduce phase workers to assign to a single-key bucket is determined based on a respective frequency of the data key assigned to the single-key bucket and a threshold level of acceptable imbalance in reduce phase worker loads across reduce phase workers.