You can also see what it looks like from the server side. nodetool proxyhistograms will show you full request latency recorded by the coordinator. nodetool cfhistograms will show you the local read latency, this is just the time it takes to read data on a replica and does not include network or wait times.
If the proxyhistograms is showing most requests running faster than your app says it's your app. Cheers ----------------- Aaron Morton Freelance Cassandra Developer New Zealand @aaronmorton http://www.thelastpickle.com On 19/01/2013, at 8:16 AM, Tyler Hobbs <ty...@datastax.com> wrote: > The fact that it's still exactly 521 seconds is very suspicious. I can't > debug your script over the mailing list, but do some sanity checks to make > sure there's not a bottleneck somewhere you don't expect. > > > On Fri, Jan 18, 2013 at 12:44 PM, Pradeep Kumar Mantha <pradeep...@gmail.com> > wrote: > Hi, > > Thanks Tyler. > > Below is the *global* connection pool I am trying to use, where the > server_list contains all the ips of 12 DataNodes I am using and > pool_size is the number of threads and I just set to timeout to 60 to > avoid connection retry errors. > > pool = pycassa.ConnectionPool('Blast', > server_list=server_list,pool_size=32,timeout=60) > > > It seems the performance is still stuck at 521 seconds.. which is 177 > seconds for cassandra-cli. > > Am I still missing something? > > thanks > Pradeep > > > > On Fri, Jan 18, 2013 at 7:12 AM, Tyler Hobbs <ty...@datastax.com> wrote: > > You just need to increase the ConnectionPool size to handle the number of > > threads you have using it concurrently. Set the pool_size kwarg to at least > > the number of threads you're using. > > > > > > On Thu, Jan 17, 2013 at 6:46 PM, Pradeep Kumar Mantha <pradeep...@gmail.com> > > wrote: > >> > >> Thanks Tyler. > >> > >> I just moved the pool and cf which store the connection pool and CF > >> information to have global scope. > >> > >> Increased the server_list values from 1 to 4. ( i think i can increase > >> them max to 12 since I have 12 data nodes ) > >> > >> when I created 8 threads using python threading package , I see the > >> below error. > >> > >> Exception in thread Thread-3: > >> Traceback (most recent call last): > >> File > >> "/usr/common/usg/python/2.7.1-20110310/lib64/python2.7/threading.py", > >> line 530, in __bootstrap_inner > >> self.run() > >> File "my_cc.py", line 20, in run > >> start_cassandra_client(self.name) > >> File "my_cc.py", line 33, in start_cassandra_client > >> cf.get(key) > >> File > >> "/global/homes/p/pmantha/mypython_repo/lib/python2.7/site-packages/pycassa/columnfamily.py", > >> line 652, in get > >> read_consistency_level or self.read_consistency_level) > >> File > >> "/global/homes/p/pmantha/mypython_repo/lib/python2.7/site-packages/pycassa/pool.py", > >> line 553, in execute > >> conn = self.get() > >> File > >> "/global/homes/p/pmantha/mypython_repo/lib/python2.7/site-packages/pycassa/pool.py", > >> line 536, in get > >> raise NoConnectionAvailable(message) > >> NoConnectionAvailable: ConnectionPool limit of size 5 reached, unable > >> to obtain connection after 30 seconds > >> > >> > >> Please have a look at the script attached.. and let me know if I need > >> to change something.. Please bear with me, if I do something terribly > >> wrong.. > >> > >> I am running the script on a 8 processor node. > >> > >> thanks > >> pradeep > >> > >> On Thu, Jan 17, 2013 at 4:18 PM, Tyler Hobbs <ty...@datastax.com> wrote: > >> > ConnectionPools and ColumnFamilies are thread-safe in pycassa, and it's > >> > best > >> > to share them across multiple threads. Of course, when you do that, > >> > make > >> > sure to make the ConnectionPool large enough to support all of the > >> > threads > >> > making queries concurrently. I'm also not sure if you're just omitting > >> > this, but pycassa's ConnectionPool will only open connections to servers > >> > you > >> > explicitly include in server_list; there's no autodiscovery of other > >> > nodes > >> > going on. > >> > > >> > Depending on your network latency, you'll top out on python performance > >> > with > >> > a fairly low number of threads due to the GIL. It's best to use > >> > multiple > >> > processes if you really want to benchmark something. > >> > > >> > > >> > On Thu, Jan 17, 2013 at 6:05 PM, Pradeep Kumar Mantha > >> > <pradeep...@gmail.com> > >> > wrote: > >> >> > >> >> Hi, > >> >> > >> >> Thanks. I would like to benchmark cassandra with our application so > >> >> that we understand the details of how the actual benchmarking is done. > >> >> Not sure, how easy it would be to integrate YCSB with our application. > >> >> > >> >> So, i am trying different client interfaces to cassandra. > >> >> > >> >> I found > >> >> > >> >> for 12 Data Nodes Cassandra cluster and 1 Client Node which run 32 > >> >> threads ( each querying X number of queries ). > >> >> > >> >> cassandra-cli took 133 seconds > >> >> pycassa took 521 seconds. > >> >> > >> >> Here is the python pycassa code used to query and passed to each > >> >> thread.... > >> >> > >> >> def start_cassandra_client(Threadname): > >> >> pool = pycassa.ConnectionPool('Blast', > >> >> server_list=['xxx.xx.xx.xx']) > >> >> cf = pycassa.ColumnFamily(pool, 'Blast_NR') > >> >> inp_file=open("pycassa_100%_query") > >> >> for key in inp_file: > >> >> key=key.strip() > >> >> cf.get(key) > >> >> > >> >> Does Java clients like Hector/Astynax help here.. I am more > >> >> comfortable with Python than Java and our existing application is also > >> >> in Python. > >> >> > >> >> thanks > >> >> pradeep > >> >> > >> >> > >> >> On Thu, Jan 17, 2013 at 2:08 PM, Edward Capriolo > >> >> <edlinuxg...@gmail.com> > >> >> wrote: > >> >> > Wow you managed to do a load test through the cassandra-cli. There > >> >> > should be > >> >> > a merit badge for that. > >> >> > > >> >> > You should use the built in stress tool or YCSB. > >> >> > > >> >> > The CLI has to do much more string conversion then a normal client > >> >> > would > >> >> > and > >> >> > it is not built for performance. You will definitely get better > >> >> > numbers > >> >> > through other means. > >> >> > > >> >> > On Thu, Jan 17, 2013 at 2:10 PM, Pradeep Kumar Mantha > >> >> > <pradeep...@gmail.com> > >> >> > wrote: > >> >> >> > >> >> >> Hi, > >> >> >> > >> >> >> I am trying to maximize execution of the number of read > >> >> >> queries/second. > >> >> >> > >> >> >> Here is my cluster configuration. > >> >> >> > >> >> >> Replication - Default > >> >> >> 12 Data Nodes. > >> >> >> 16 Client Nodes - used for querying. > >> >> >> > >> >> >> Each client node executes 32 threads - each thread executes 76896 > >> >> >> read > >> >> >> queries using cassandra-cli tool. > >> >> >> i.e all the read queries are stored in a file and that file > >> >> >> is > >> >> >> given to cassandra-cli tool ( using -f option ) which is executed by > >> >> >> a > >> >> >> thread. > >> >> >> so, total number of queries for 16 client Nodes is 16 * 32 * 76896. > >> >> >> > >> >> >> The read queries on each client node submitted at the same time. The > >> >> >> time taken for 16 * 32 * 76896 read queries is nearly 742 seconds - > >> >> >> which is nearly 53k transactions/second. > >> >> >> > >> >> >> I would like to know if there is any other way/tool through which I > >> >> >> can improve the number of transactions/second. > >> >> >> Is the performance affected by cassandra-cli tool? > >> >> >> > >> >> >> thanks > >> >> >> pradeep > >> >> > > >> >> > > >> > > >> > > >> > > >> > > >> > -- > >> > Tyler Hobbs > >> > DataStax > > > > > > > > > > -- > > Tyler Hobbs > > DataStax > > > > -- > Tyler Hobbs > DataStax
