incubator-cassandra-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Jack Krupansky" <j...@basetechnology.com>
Subject Re: memtable mem usage off by 10?
Date Wed, 04 Jun 2014 14:54:47 GMT
And sorry that the doc confused you as well!

-- Jack Krupansky

From: Idrén, Johan 
Sent: Wednesday, June 4, 2014 10:51 AM
To: user@cassandra.apache.org 
Subject: Re: memtable mem usage off by 10?

I wasn’t supplying it, I was assuming it was using the default. It does not exist in my
config file. Sorry for the confusion.




From: Benedict Elliott Smith <belliottsmith@datastax.com>
Reply-To: "user@cassandra.apache.org" <user@cassandra.apache.org>
Date: Wednesday 4 June 2014 16:36
To: "user@cassandra.apache.org" <user@cassandra.apache.org>
Subject: Re: memtable mem usage off by 10?


  Oh, well ok that explains why I'm not seeing a flush at 750MB. Sorry, I was going by the
documentation. It claims that the property is around in 2.0.
But something else is wrong, as Cassandra will crash if you supply an invalid property, implying
it's not sourcing the config file you're using.


I'm afraid I don't have the context for why it was removed, but it happened as part of the
2.0 release.



On 4 June 2014 13:59, Jack Krupansky <jack@basetechnology.com> wrote:

  Yeah, it is in the doc:
  http://www.datastax.com/documentation/cassandra/2.0/cassandra/configuration/configCassandra_yaml_r.html

  And I don’t find a Jira issue mentioning it being removed, so... what’s the full story
there?!

  -- Jack Krupansky

  From: Idrén, Johan 
  Sent: Wednesday, June 4, 2014 8:26 AM
  To: user@cassandra.apache.org 
  Subject: RE: memtable mem usage off by 10?

  Oh, well ok that explains why I'm not seeing a flush at 750MB. Sorry, I was going by the
documentation. It claims that the property is around in 2.0.




  If we skip that, part of my reply still makes sense:




  Having memtable_total_size_in_mb set to 20480, memtables are flushed at a reported value
of ~2GB. 




  With a constant overhead of ~10x, as suggested, this would mean that it used 20GB, which
is 2x the size of the heap.




  That shouldn't work. According to the OS, cassandra doesn't use more than ~11-12GB.





------------------------------------------------------------------------------

  From: Benedict Elliott Smith <belliottsmith@datastax.com>
  Sent: Wednesday, June 4, 2014 2:07 PM
  To: user@cassandra.apache.org
  Subject: Re: memtable mem usage off by 10? 

  I'm confused: there is no flush_largest_memtables_at property in C* 2.0?



  On 4 June 2014 12:55, Idrén, Johan <Johan.Idren@dice.se> wrote:

    Ok, so the overhead is a constant modifier, right.




    The 3x I arrived at with the following assumptions:




    heap is 10GB


    Default memory for memtable usage is 1/4 of heap in c* 2.0


    max memory used for memtables is 2,5GB (10/4)
    flush_largest_memtables_at is 0.75


    flush largest memtables when memtables use 7,5GB (3/4 of heap, 3x of the default)




    With an overhead of 10x, it makes sense that my memtable is flushed when the jmx data
says it is at ~250MB, ie 2,5GB, ie 1/4 of the heap




    After I've set the memtable_total_size_in_mb to a value larger than 7,5GB, it should still
not go over 7,5GB on account of flush_largest_memtables_at, 3/4 the heap




    So I would expect to see memtables flushed to disk after they're being reportedly at around
750MB.




    Having memtable_total_size_in_mb set to 20480, memtables are flushed at a reported value
of ~2GB. 




    With a constant overhead, this would mean that it used 20GB, which is 2x the size of the
heap, instead of 3/4 of the heap as it should be if flush_largest_memtables_at was being respected.




    This shouldn't be possible.





----------------------------------------------------------------------------

    From: Benedict Elliott Smith <belliottsmith@datastax.com>

    Sent: Wednesday, June 4, 2014 1:19 PM 

    To: user@cassandra.apache.org
    Subject: Re: memtable mem usage off by 10?

    Unfortunately it looks like the heap utilisation of memtables was not exposed in earlier
versions, because they only maintained an estimate. 

    The overhead scales linearly with the amount of data in your memtables (assuming the size
of each cell is approx. constant). 

    flush_largest_memtables_at is an independent setting to memtable_total_space_in_mb, and
generally has little effect. Ordinarily sstable flushes are triggered by hitting the memtable_total_space_in_mb
limit. I'm afraid I don't follow where your 3x comes from?




    On 4 June 2014 12:04, Idrén, Johan <Johan.Idren@dice.se> wrote:

      Aha, ok. Thanks.




      Trying to understand what my cluster is doing:




      cassandra.db.memtable_data_size only gets me the actual data but not the memtable heap
memory usage. Is there a way to check for heap memory usage?





      I would expect to hit the flush_largest_memtables_at value, and this would be what causes
the memtable flush to sstable then? By default 0.75?




      Then I would expect the amount of memory to be used to be maximum ~3x of what I was
seeing when I hadn't set memtable_total_space_in_mb (1/4 by default, max 3/4 before a flush),
instead of close to 10x (250mb vs 2gb).


      This is of course assuming that the overhead scales linearly with the amount of data
in my table, we're using one table with three cells in this case. If it hardly increases at
all, then I'll give up I guess :)

      At least until 2.1.0 comes out and I can compare.




      BR

      Johan




--------------------------------------------------------------------------

      From: Benedict Elliott Smith <belliottsmith@datastax.com>

      Sent: Wednesday, June 4, 2014 12:33 PM 

      To: user@cassandra.apache.org
      Subject: Re: memtable mem usage off by 10?

      These measurements tell you the amount of user data stored in the memtables, not the
amount of heap used to store it, so the same applies.



      On 4 June 2014 11:04, Idrén, Johan <Johan.Idren@dice.se> wrote:

        I'm not measuring memtable size by looking at the sstables on disk, no. I'm looking
through the JMX data. So I would believe (or hope) that I'm getting relevant data. 




        If I have a heap of 10GB and set the memtable usage to 20GB, I would expect to hit
other problems, but I'm not seeing memory usage over 10GB for the heap, and the machine (which
has ~30gb of memory) is showing ~10GB free, with ~12GB used by cassandra, the rest in caches.




        Reading 8k rows/s, writing 2k rows/s on a 3 node cluster. So it's not idling.




        BR


        Johan





------------------------------------------------------------------------

        From: Benedict Elliott Smith <belliottsmith@datastax.com>
        Sent: Wednesday, June 4, 2014 11:56 AM
        To: user@cassandra.apache.org
        Subject: Re: memtable mem usage off by 10? 

        If you are storing small values in your columns, the object overhead is very substantial.
So what is 400Mb on disk may well be 4Gb in memtables, so if you are measuring the memtable
size by the resulting sstable size, you are not getting an accurate picture. This overhead
has been reduced by about 90% in the upcoming 2.1 release, through tickets 6271, 6689 and
6694.



        On 4 June 2014 10:49, Idrén, Johan <Johan.Idren@dice.se> wrote:

          Hi,




          I'm seeing some strange behavior of the memtables, both in 1.2.13 and 2.0.7, basically
it looks like it's using 10x less memory than it should based on the documentation and options.




          10GB heap for both clusters.


          1.2.x should use 1/3 of the heap for memtables, but it uses max ~300mb before flushing




          2.0.7, same but 1/4 and ~250mb




          In the 2.0.7 cluster I set the memtable_total_space_in_mb to 4096, which then allowed
cassandra to use up to ~400mb for memtables...




          I'm now running with 20480 for memtable_total_space_in_mb and cassandra is using
~2GB for memtables.




          Soo, off by 10 somewhere? Has anyone else seen this? Can't find a JIRA for any bug
connected to this.


          java 1.7.0_55, JNA 4.1.0 (for the 2.0 cluster)




          BR


          Johan







Mime
View raw message