eXtremeDB cluster design considerations

In the cluster environment, each WRITE transaction requires onenetwork round-trip: the initiator sends transaction data and waits forresponses from all other nodes. Thus the duration of a transaction cannot beless than the network latency (or 'ping time'). For Ethernet the typical WRITEtransaction rate is 5-7K TPS for one thread.

 

Some common methods to increase the overall cluster performanceare:

-         Combine several small WRITE-transactions into one largetransaction so that, instead of many network round-trips, there will be onlyone.

-         Execute WRITE-transactions in several threads if possible. Thisway each thread sends and receives network packages independently, so that youcould obtain near-linear dependency of performance from the number of threads.

-         If using several write threads, it's worth using the transactionwindow feature. Try setting the cl_params.window.length to about one half ofthe number of 'active' write threads. 

Note that in a cluster, all nodes can write atthe same time.  Thus, given your example of 40K TPS, two nodescould achieve 80K TPS in the aggregate, assuming no conflicts.  Note alsothat read-only transactions don’t cause any network traffic, so every nodemultiplies the number of read-only transactions that are possible.