Some comments inline
On 7/29/14, 1:07 PM, Pelton, Aaron A. wrote:
Hi All,
I am new to Accumulo and I apologize if the answers to my questions are
already posted somewhere. I’ve done a fair amount of googling and poking
around the manuals etc.
I am just doing a simple test with two machines, one producing about 600
threads on the network to stream simultaneous writes to a rest service,
and the other producing about 300 threads on the network to perform
simultaneous queries to a rest service. The rest service has Accumulo
API calls in it to write out and query data.
I have inherited the following configuration
-Squirrel Bundle distribution of Accumulo 1.5.0
-1 Master machine to start and stop Accumulo services on
-12 data nodes running tservers. The first three of these also running
the zookeeper instances. And, nodes 4-6 running tracers.
I have noticed the following issues with configuration and changed them
as follows
-Changed swapiness to 0 on all nodes
-Was getting OutOfMemoryExceptions after the above still, and after
running test for long duration. Thus, increased Java Heap size from 1g
to 4g, which is still far below the physical ram on the nodes.
-Increased java heap from 1g to 2g on master node
-I also increased the following properties
o <property>
o <name>tserver.memory.maps.max</name>
o <value>2G</value>
o </property>
o
o <property>
o <name>tserver.cache.data.size</name>
o <value>512M</value>
o </property>
o
o <property>
o <name>tserver.cache.index.size</name>
o <value>512M</value>
o </property>
-Changed the ulimit for virtual memory to unlimited
-Changed the ulimit for files opened to 65536
-Changed the ulimit for max user processes to 1024
These all look good. Just keep in mind that tserver.cache.data.size and
tserver.cache.index.size will be on the JVM heap while
tserver.memory.maps.max is off heap (assuming you're using the native
maps which you very well should be -- I assume Sqrrl's distro set this
up for you)
-A tomcat instance with a server socket accepting up to 1,000 threads /
user connections to a rest service that eventually makes a read / write
out to an Accumulo connector instance.
-Changed the zookeeper connection limit max to 0 since this is just a
test environment
-Noticed that code I had inherited didn’t have close calls on the
scanner objects in the rest service b/c it was originally designed for
Accumulo 1.4 in which there wasn’t such an API.
Scanners can clean up after themselves, whereas BatchScanners don't. A
close method was added to ScannerBase (the parent class of Scanner and
BatchScanner) to let you seamlessly swap out a Scanner with a
BatchScanner (and vice versa) while not leaking any resources. In short,
you can call Scanner#close, but it's just a no-op.
-This may be wrong, but in an effort to see my ~900 connections
simultaneously get as much access to db writes/reads for servicing, I
up’d some thread counts for
o <property>
o <name>tserver.server.threads.minimum</name>
o <value>75</value>
o </property>
o
o <property>
o <name>master.server.threads.minimum</name>
o <value>300</value>
o </property>
I have a couple of problems to note:
1.Ingest speeds seem kinda slow. I would anticipate network overhead but
not enough to reduce writes to 125 records / sec when each record is
only a few kB.
What do you actually do when you receive an HTTP request to write to
Accumulo. It sounds like you're reading data and then writing? Is each
HTTP request creating its own BatchWriter? More insight to what a
"write" looks like in your system (in terms of Accumulo API calls) would
help us make recommendations about more efficient things you can do.
a.I believe this is due to the fact that I’m only seeing one tserver
primarily active at ingesting, with one tbalet in particular for the
table receiving the bulk of the data.
b.I have added pre-splits upon table creation for each letter of the
alphabet, plus the digits 0-9. As this is a test with a simple loop
creating ID values, I throw 2 alpha chars randomly in front of the
generated number in my loop and use that as the ID to distribute
hopefully the IDs across tablets for this table. A record ID ingested
might look like “bk1234:8876”, whereby it has random 2 chars, orig ID
value, colon, and a timestamp. Sample pre-splitting: (Granted the array
could be constructed more gracefully, but for a quick test, meh).
*try*
{
conn.tableOperations().create(/TABLE_NAME/);
*final*SortedSet<Text> sortedSplits = *new*TreeSet<Text>();
*for*(String binPrefix : *new*String[] { "a", "b", "c", "d", "e", "f",
"g", "h", "i", "j", "k", "l", "m",
"n", "o", "p", "q", "r", "s", "t", "u", "v", "w", "x", "y", "z", "1",
"2", "3", "4", "5", "6", "7",
"8", "9", "0"})
{
sortedSplits.add(*new*Text(binPrefix));
}
conn.tableOperations().addSplits(/TABLE_NAME/, sortedSplits);
}
*catch*(TableExistsException | TableNotFoundException exception)
{
/LOGGER/.warn("Could not create table or sorted splits", exception);
}
Good, pre-splitting your table should help with random data, but if
you're only writing data to one tablet, you're stuck (very similar to
hot-spotting reducers in MapReduce jobs).
2.Tservers running on the data node halt after about 4 hours in of
processing. I’m attempting to ingest into the billions, hopefully
trillions of records range. Generally it is the ones that aren’t under
load in the beginning, until finally the one that is handling the bulk
of the load crashes typically last. In the beginning, I noticed in the
tserver logs the OutOfMemoryException, but haven’t seen that in the past
few runs after the memory adjustments. In fact the tserver log doesn’t
say anything about why it stopped. Also didn’t notice anything unusual
in the zookeeper log other than the occasional CancelledKeyException.
Make sure you check both the tserver_hostname.debug.log,
tserver_hostname.out and tserver_hostname.err files. OOMEs sometimes
don't make it to the log file because of the JVM tearing down. You
should be able to find something as to why the tserver stopped.
3.Lastly can anyone approximate with the 12 nodes that I have, what kind
of ingest speed should I see if things were configured correctly in
number of records per second based on small record sizes of a few kB.
And, is anything obviously wrong with the configurations mentioned above
that would improve throughput?
Generally, a "normal" machine will be able to do ingest of about 200k
records at 150bytes for ~30MB/s.
You might also want to try increasing tserver.mutation.queue.max to 1M
in accumulo-site.xml (restart required). You can find some extra
information about that on the releases notes:
http://accumulo.apache.org/release_notes/1.5.1.html#known-issues. Not
sure if Sqrrl's distribution has done this already for you.
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
--
Sincerely,
Aaron Pelton
