Skip to content

Comment on Nobody ever got fired for buying a clusterparent

Comments

Yes, section 3.4: "Heap size - By default each Hadoop map and reduce task is run in a JVM with a 200 MB heap within which they allocate buffers for in-memory data. When the buffers are full, data is spilled to storage, adding overheads. We note that 200 MB per task leaves substantial amounts of memory unused on modern servers. By increasing the heap size for each JVM (and hence the working memory for each task), we improve performance. However too large a heap size causes garbage collection overheads, and wastes memory that could be used for other purposes (such as a RAMdisk). For the scale-out configurations, we found the optimal heap size for each job through trial and error. For the scale-up configuration we set a heap size of 4 GB per mapper/reducer task (where the maximum number of tasks is set to the number of processors) for all jobs."

AboutSource Built by g1lg1l

Hackerly is an independent reader for Hacker News, built on the public HN API. Not affiliated with Y Combinator.