tree 93a1ac914dab21c4e08c630178aa6f8235f2d6f6
parent 117fc388bff2a754be081eae7667627f84f1b33c
author Sailesh Mukil <sailesh@cloudera.com> 1496173813 +0000
committer Taras Bobrovytsky <tbobrovytsky@cloudera.com> 1496364738 -0700

IMPALA-5378: Disk IO manager needs to understand ADLS

The Disk IO Manager had customized support for S3 and remote HDFS that
allows for these to use a separate queue and have a customized number
of IO threads. ADLS did not have this support.

Based on the code in DiskIoMgr::Init and DiskIoMgr::AssignQueue, IOs
for ADLS were previously put in local disk queues. Since local disks
are considered rotational unless we can confirm otherwise by looking at
the /sys filesystem, this means that THREADS_PER_ROTATIONAL_DISK=1 was
being applied as the thread count.

This patch adds customized support for ADLS, similar to how it was done
for S3. We set 16 threads as the default number of IO threads for ADLS.
For smaller clusters, setting a higher number like 64 would work better.
We keep the thread count to a lower default of 16 since there is an
undocumented concurrency limit for clusters, which is around 500-700
connections, which means we would hurt node level parallelism if we
have higher thread level parallelism, for larger clusters.

We also set the default maximum chunk size for ADLS as 128k. This is due
to the fact that direct reads aren't supported for ADLS, which means that
the JNI array allocation and the memcpy adds significant overhead for
larger buffers. 128k was chosen emperically for S3 for the same reason.
Since this reason also holds for ADLS, we keep the same value. A new
flag called FLAGS_adls_read_chunk_size is used to control this value.

TODO: Settle on a buffer size with the most optimal buffer size
emperically.

Change-Id: I067f053fec941e3631610c5cc89a384f257ba906
Reviewed-on: http://gerrit.cloudera.org:8080/7033
Reviewed-by: Sailesh Mukil <sailesh@cloudera.com>
Tested-by: Impala Public Jenkins