Hi Aaron, Thank you,and your explanation makes sense. At the time, I thought having 1GB of row cache on each node was plenty enough, because there was an aggregated 6GB cache, but you are right, with each row in 10's of MBs, some of the nodes can go into a constant load and evict cycle and would have negative effects on the performance. I will try as you suggested to 1.) reduce the requested entry set, and 2.) increase the row cache size and see if they get better hits, and also do 3) by reversing the requested entry list in alternate runs.
Our data space has close to 3 million rows, but we haven't gotten enough usage statistics to know what rows are hot. Does this mean we should not enable row caches until we are absolutely sure about what's hot (I think there is a reason why row caches are disabled by default) ? It also seems from my test that OS page cache works much better, but it could be that OS page cache can utilize all the available memory so it is essentially larger -- I guess I will find out by doing 2.) above. best, -- Y. On Tue, Dec 4, 2012 at 4:47 AM, aaron morton <aa...@thelastpickle.com>wrote: > > Row Cache : size 1072651974 (bytes), capacity 1073741824 (bytes), > 0 hits, 2576 requests, NaN recent hit rate, 0 save period in seconds > > So the cache is pretty much full, there is only 1 MB free. > > There were 2,576 read requests that tried to get a row from the cache. > Zero of those had a hit. If you have 6 nodes and RF 2, each node has one > third of the data in the cluster (from the effective ownership info). So > depending on the read workload the number of read requests on each node may > be different. > > What I think is happening is reads are populating the row cache, then > subsequent reads are evicting items from the row cache before you get back > to reading the original rows. So if you read rows 1 to 5, they are put in > the cache, when you read rows 6 to 10 they are put in and evict rows 1 to > 5. Then you read rows 1 to 5 again they are not in the cache. > > Try testing with a lower number of hot rows, and/or a bigger row cache. > > But to be honest, with rows in the 10's of MB you will probably only get > good cache performance with a small set of hot rows. > > Hope that helps. > > > > ----------------- > Aaron Morton > Freelance Cassandra Developer > New Zealand > > @aaronmorton > http://www.thelastpickle.com > > On 1/12/2012, at 5:11 AM, Yiming Sun <yiming....@gmail.com> wrote: > > > Does anyone have any comments/suggestions for me regarding this? Thanks > > > > > > I am trying to understand some strange behavior of cassandra row cache. > We have a 6-node Cassandra cluster in a single data center on 2 racks, and > the neighboring nodes on the ring are from alternative racks. Each node > has 1GB row cache, with key cache disabled. The cluster uses > PropertyFileSnitch, and the ColumnFamily I fetch from uses > NetworkTopologyStrategy, with replication factor of 2. My client code uses > Hector to fetch a fixed set of rows from cassandra > > > > What I don't quite understand is even after I ran the client code > several times, there are always some nodes with 0 row cache hits, despite > that the row cache from all nodes are filled and all nodes receive requests. > > > > Which nodes have 0 hits seem to be strongly related to the following: > > > > - the set of row keys to fetch > > - the order of the set of row keys to fetch > > - the list of hosts passed to Hector's CassandraHostConfigurator > > - the order of the list of hosts passed to Hector > > > > Can someone shed some lights on how exactly the row cache works and > hopefully also explain the behavior I have been seeing? I thought if the > fixed set of the rows keys are the only thing I am fetching (each row > should be on the order of 10's of MBs, no more than 100MB), and each node > gets requests, and its row cache is filled, there's gotta be some hits. > Apparent this is not the case. Thanks. > > > > cluster information: > > > > Address DC Rack Status State Load > Effective-Ownership Token > > > 141784319550391026443072753096570088105 > > x.x.x.1 DC1 r1 Up Normal 587.46 GB 33.33% > 0 > > x.x.x.2 DC1 r2 Up Normal 591.21 GB 33.33% > 28356863910078205288614550619314017621 > > x.x.x.3 DC1 r1 Up Normal 594.97 GB 33.33% > 56713727820156410577229101238628035242 > > x.x.x.4 DC1 r2 Up Normal 587.15 GB 33.33% > 85070591730234615865843651857942052863 > > x.x.x.5 DC1 r1 Up Normal 590.26 GB 33.33% > 113427455640312821154458202477256070484 > > x.x.x.6 DC1 r2 Up Normal 583.21 GB 33.33% > 141784319550391026443072753096570088105 > > > > > > [user@node]$ ./checkinfo.sh > > *************** x.x.x.4 > > Token : 85070591730234615865843651857942052863 > > Gossip active : true > > Thrift active : true > > Load : 587.15 GB > > Generation No : 1354074048 > > Uptime (seconds) : 36957 > > Heap Memory (MB) : 2027.29 / 3948.00 > > Data Center : DC1 > > Rack : r2 > > Exceptions : 0 > > > > Key Cache : size 0 (bytes), capacity 0 (bytes), 0 hits, 0 > requests, NaN recent hit rate, 14400 save period in seconds > > Row Cache : size 1072651974 (bytes), capacity 1073741824 (bytes), > 0 hits, 2576 requests, NaN recent hit rate, 0 save period in seconds > > > > *************** x.x.x.6 > > Token : 141784319550391026443072753096570088105 > > Gossip active : true > > Thrift active : true > > Load : 583.21 GB > > Generation No : 1354074461 > > Uptime (seconds) : 36535 > > Heap Memory (MB) : 828.71 / 3948.00 > > Data Center : DC1 > > Rack : r2 > > Exceptions : 0 > > > > Key Cache : size 0 (bytes), capacity 0 (bytes), 0 hits, 0 > requests, NaN recent hit rate, 14400 save period in seconds > > Row Cache : size 1072602906 (bytes), capacity 1073741824 (bytes), > 0 hits, 3194 requests, NaN recent hit rate, 0 save period in seconds > > > > > >