| On a shared-nothing architecture, one or more disks are connectedto each processor, depending on machineÕs (parallel) I/O architecture, via direct connecting or some I/O channels. |
| The shared-nothing does not focus on the physical I/O connection architectures, rather, it refers tohow the data is partitioned on the disk arrays and how the data are movedinto the procesor buffers for the parallel query processing. |
| In this architecture, data locations determines where the data will be processed and how data is shared by processors other than its local processor (ie. the processor which has the local disk holding the data if direct connection like DB2 on SP2, or the shortest path from disks to the process if other I/O architectures like Oracle on ncube). |
| Unlike shared-disk where data may be shared via the interconnection network when it is first readfrom a disk to a remote processor (buffer), if the data local to processorA is required by a remote processor B to perfom some parallel query processing, A sends the data to B via the communication network. |
| That is the only shared media for shared-nothing is the communication network. |
| In most case, data placement has determined how and where (even when) a parallel query processing is decomposed or partitioned. |
| Data (or I/O) load balance determines the CPU load balance in the system. |
| This is built in the query decomposer. eg. |
| if Processor A gets 10% data local to its disk, while B gets remaining 90%. |
| Then in a shared-nothing system, A spends 10% CPU and then is idle while B takes most of CPU time. |
| On the otherhand for a shared-disk system, A and B will process similiar amount of data with some overhead of sending data from B to A. |