accumulo-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Pelton, Aaron A." <Aaron.Pel...@gd-ais.com>
Subject RE: Request for Configuration Help for basic test. Tservers dying and only one tablet being used
Date Tue, 29 Jul 2014 20:53:03 GMT
I agree with the idea of pooling the writers.

As for the discussion of the keys. I get what you are saying with choosing better keys for
distribution based on frequency of the chars in the English language. But, for this test I'm
just using apache RandomStringUtils to create a 2 char random alpha sequence to prepend, so
it should be a moderately distributed sampling of chars. However, let me emphasize that I
mean I'm seeing 1 tablet getting millions of entries in it, compared to the remaining 35 tablets
having no entries or just like 1k. To me that says something isn't right.


-----Original Message-----
From: Josh Elser [mailto:josh.elser@gmail.com] 
Sent: Tuesday, July 29, 2014 4:20 PM
To: user@accumulo.apache.org
Subject: Re: Request for Configuration Help for basic test. Tservers dying and only one tablet
being used

On 7/29/14, 3:20 PM, Pelton, Aaron A. wrote:
> To followup to two of your statements/questions:
>
> 1. 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).
>
> - OK so its good that the data is presplitting, but maybe this is conceptually something
that I'm not grasping about accumulo yet, but I thought specifying the pre-splits is what
causes the table to span multiple tablets on the various tserver initially.  However, the
core of the data appears to be in one specific tablet on on tserver. Each tserver appears
to have a few tablets allocated to it for the table I'm working out of. So, I'm confused as
to how to get the data to write to more than just the one tablet/partition.  I would almost
think my keys I specified aren't being matched correctly against incoming data then?

No, it sounds like you have the idea correctly. Many tablets make up a table, the split points
for a table are what defines those tablet boundaries. Consider you have a table where the
rowID are English words (http://en.wikipedia.org/wiki/Letter_frequency#Relative_frequencies_of_the_first_letters_of_a_word_in_the_English_language).

If you split your table on each letter (a-z), you would still see much more activity to the
tablets which host words starting with 'a', 't', and 's' because you have significantly more
data being ingested into those tablets.

When designing a table (specifically the rowID of the key), it's desirable to try to make
the rowID as distributed as possible across the entire table. This helps ensure even processing
across all of your nodes. Does that make sense?

> 2. 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.
>
> Yes each http request gets its own reference to a writer or scanner, which is closed
when thre result is returned from the http request.  There are two rest services. One transforms
the data and preforms some indexes based on it and then sends both data and index to a BatchWriter.
The sample code for the data being written is below. The indexes being written are similar
but use different family and qualifier values.
>
>          Text rowId = new Text(id + ":" + time);
>          Text fam = new Text(COLUMN_FAMILY_KLV);
>          Text qual = new Text("");
>          Value val = new Value(data.getBytes());
>
>          Mutation mut = new Mutation(rowId);
>          mut.put(fam, qual, val);
>
>          long memBuf = 1_000_000L;
>          long timeout = 1000L;
>          int numThreads = 10;
>
>          BatchWriter writer = null;
>          try
>          {
>              writer = conn.createBatchWriter(TABLE_NAME, memBuf, timeout, numThreads);
>              writer.addMutation(mut);
>          }
>          catch (Exception x)
>          {
>              // x.printStackTrace();
>              logger.error(x.toString(), x);
>              result = "ERROR";
>          }
>          finally
>          {
>              try
>              {
>                  if (writer != null)
>                  {
>                      writer.close();
>                  }
>              }
>              catch (Exception x)
>              {
>                  // x.printStackTrace();
>                  logger.error(x.toString(), x);
>                  result = "ERROR";
>              }
>          }

You could try to make a threadpool for BatchWriters instead of creating a new one for each
HTTP thread. This might help amortize the RPC cost by sending more than one mutation at a
time (the BatchWriter should be thread safe in this regard). You then just want to call flush()
instead of closing the BatchWriter.

I remember seeing that there are some optimizations within the BatchWriter to write a single
Mutation, but if you're really trying to saturate your system, using fewer BatchWriters would
likely help you realize more throughput.

> At the beginning of the test, a known subset of control data range is created and uploaded.
For the duration of the heart of the test while ongoing writes occur, queries upon data in
that control range are performed.  The rest service that handles the read eventually hits
this:
>
>          ArrayList<String> latlons = new ArrayList<String>();
>          Authorizations auths = new Authorizations();
>
>          Scanner scan = null;
>          try
>          {
>              scan = conn.createScanner(TABLE_NAME, auths);
>              scan.setRange(new Range(id + ":0", id + "::")); // all times
>              scan.fetchColumnFamily(new Text(COLUMN_FAMILY_KLV));
>
>              for (Map.Entry<Key, Value> e : scan)
>              {
>                  // do stuff with e
>              }
>          }
>          catch (TableNotFoundException x)
>          {
>              LOGGER.fatal("The table " + TABLE_NAME + " could not be found.", x);
>          }
>          finally
>          {
>              if (scan != null)
>              {
>                  scan.close();
>              }
>          }
>
> -----Original Message-----
> From: Josh Elser [mailto:josh.elser@gmail.com]
> Sent: Tuesday, July 29, 2014 1:43 PM
> To: user@accumulo.apache.org
> Subject: Re: Request for Configuration Help for basic test. Tservers 
> dying and only one tablet being used
>
> 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
>>

Mime
View raw message