IMO, I see this at home on my computer no matter what memory settings I
use. I've become pretty accustomed to flat out ignoring it...
As for heap management, there are two big paths here: with "native maps"
and without. When you write data to Accumulo, it goes to two places: 1)
Write-ahead log and 2) Memory maps. The WAL ensures that if you have
writes in memory on a server that dies, that you don't lose data. The
memory maps give you much faster ingest over trying to write into a
sorted file.
1) Native maps (aka c++ code over JNI)
This memory allocation, controlled by tserver.memory.maps.max in
accumulo-site.xml, is "off heap" memory. It is not limited by the JVM
heap limits you specify in ACCUMULO_TSERVER_OPTS in accumulo-env.sh. As
such, you need to make sure that you don't over-allocate memory usage on
your node (tserver.memory.maps.max + JVM Xmx + fudge-factor < total
available memory).
2) Non-native (in JVM)
This serves the same purpose as #1 but is in JVM heap as opposed to off
heap. Ingest will be slower and JVM gc will likely be a bigger issue
than using the native maps. This does make the JVM sizing a little more
straightforward: JVM Xmx + fudge-factor < total available memory (but
math is pretty easy).
Assuming you use the native maps, lets break down what you see in JVM heap.
1) Index block cache
Each RFile (backing file for tablets in Accumulo), has an multi-level
index structure which lets you efficiently find the data in that file.
Accumulo provides the ability to cache this index information instead of
reading and deserializing from disk every time. Controlled by
tserver.cache.index.size.
2) Data block cache
Similar to #1 except it's for the actual blocks of data in that RFile
(the key-value pairs) instead of just the index structure. Controlled by
tserver.cache.data.size. This can give you some benefit over having to
hit a (potentially, remote) datanode every time you perform a read in a
read-heavy environment.
3) "The rest"
Consider this the rest of the things that the tabletserver does.
"hosting" its tablets (each tablet has a collection of files in hdfs),
scansessions running against those hosted tablets (the iterator stack
that is created to perform a "read"). Compression (de)allocators for
Hadoop (assuming you're using GZIP). Various internal buffers for
caching. Connection management information (thrift and hadoop
connections). I'm probably missing more things, too.
On 11/12/13, 12:32 PM, Terry P. wrote:
On an Accumulo 1.4.2 I've gotten "[tabletserver.TabletServer] WARN:
Running low on memory" 5 times in the last two days on just one of my 6
datanodes. That datanode is hosting ~30% of the data, as 2 datanodes had
dropped from the cluster due to a network issue some time ago and hasn't
entirely rebalanced. Current volume is only 140 million.
Ingest rates has been pretty constant at a light 200 per second.
Not knowing how Accumulo uses its java heap space, I opted to start with
a stock memory config and used the 3GB example config files, which I see
allocates only 1GB to the TabletServer. The server has 24GB RAM and
currently is using only 10GB total between Accumulo and HDFS, so there's
plenty of free memory to spare.
Is ACCUMULO_TSERVER_OPTS in accumulo-env.sh the tunable I should target
to alleviate these warnings?
Unfortunately being ops-configured nodes, no JDKs are installed nor is
it a possibility to do so in order to monitor the JVM itself for better
information.
Thanks in advance,
Terry
