[rxin]

No it doesn't. The old record used 2100 nodes so the entire data actually fit in memory. There shouldn't be much seek happening even in the MR 2100 case. In Spark's case, the data actually doesn't fit in memory.

Also this was primary network bound. The old record had 2100 nodes with 10Gbps network.

[discardorama]

As per the specification of this test, the data has to be committed on disk before it is considered sorted. So even if it all fits in memory, it has to be on disk before the end.

So you have 100TB of disk read, followed by 100TB of disk write, all on HDDs. That's about 100GB/node; and since Hadoop nodes are typically in RAID-6, each write has an associated read and write too.

This does not even include the intermediate files, which (depending on how the kernel parameters have been set), could have been written on disk. Typical dirty_background_ratio is 10; so after 6GB of dirty pages, pdflush will kick in and start writing to the spinning disk.

[rxin]

Yes, but the final data is sequential only. We were discussing about random access, which only applies to the intermediate shuffle file.

Maybe you can email me offline. I can tell you more about the setup and how Spark / MapReduce works w.r.t. to it.

[lmeyerov]

Ah, so you're saying as this is network bound, you want to launch much fewer wimpy nodes in exchange for a few bigger ones (those w/ SSDs -- over 1PB worth!). If it was memory-bound, you wouldn't care how they're spread as long as it was still in-memory.

Another way of looking at this is performance per watt or dollar. The r3.2 has 60GB, so comparing to that, Spark cost the same ~$1.4K while giving a 3X speedup. (Or on host that charges per minute, it'd be the same performance at 3X cheaper.)

This is me not knowing the space: would MR (or more modern things like Tez) perform worse on this HW setup, or is this a reminder that hardware/config tuning matters?

[rxin]

Yes absolutely. If I can get a single machine with 200TB of SSDs, that'd have been great :)

But as soon as we have more than 1 node, then having more nodes is better. We can actually demonstrate this quantitatively. We are required to replicate the output, which means 100 TB of data would generate 100 TB of network for replication, and 100 * (N-1)/N TB of network for shuffle, where N = num nodes. That is the overall 200 - 100/N in network. Assuming each node has 1GB/s network, then you'd need (200 - 100/N) / N seconds just to transfer the data across network, i.e. the optimal number of nodes is the one that gives you the lowest (200/N - 100/N^2).

The problem is with last year's MR run was that it wasn't saturating the network at all. It had roughly 1.5GB/s/node HDDs, and the overall network throughput was probably around 20MB/s/node when they were using 10Gbps network (I'm assuming only half of the time is doing network. If they were doing network the full time, then network throughput was at ~10MB/s/node).

Had we been using the same HDDs last year's entry had, our map phase would slow down by about 2x, and the reduce phase shouldn't change. This would mean a total run time of less than 30 mins on ~200 nodes. Still way better.