sweiger@sequent.UUCP (Mark Sweiger) (04/13/90)
In article <1531@mountn.dec.com> wallis@labc.dec.com (Barry L. Wallis) writes: > > >Yo Patrick, have you used Oracle in an update intensive application on an SMP >machine or in a cluster. The single write server architecture is definitely a >drawback in that situation. I've seen known of Oracle lose benchmarks in that >situation. The above is not correct. Version 6 of Oracle allows multiple asychronous page cleaner processes on symmetric multiprocessors like a Sequent or a Pyramid, which eliminates the problem stated above. In general, more is better in symmetric multiprocessing, and this certainly applies to the RDBMS process architecture. Anywhere there is one of anything is a potential bottleneck. Products which allow multiple client processes, multiple server processes, and multiple asynchronous daemon processes tend to perform well on SMPs. For example, we have seen TP1 benchmark results vary from 20 to 200+ TPS on identical Sequent configurations, the only variant being whether the tested RDBMS product permitted only one server process (low TPS), vs. products which allowed multiple server processes (high TPS). If you only have one server process, you can only effectively use one of the many available CPUs to perform server-side functions, which constrains throughput. All of the Big 5 RDMBS vendors (Oracle, Informix, Sybase, Ingres, and Unify) have or will soon have unconstrained multiple process architectures for SMP machines, diminishing differences in performance caused by process-level bottlenecks. The next, and perhaps more interesting, SMP performance battles will be fought by parallelizing intra- process algorithms (parallel scan, parallel join, parallel execution of query subtrees, parallel sort, etc.). The net performance gain in this area can be even more pronounced; prototypes have shown 100 to 1 or better gains in certain query execution times. It is interesting to note that process-level and algorithmic parallelization of RDBMS architectures has shown impressive performance gains on Cube-like hardware architectures and cluster hardware architectures, not just SMPs. It is a powerful concept. -------------------------------------------------------- Mark Sweiger Sequent Computer Systems Database Software Engineer 15450 SW Koll Parkway Beaverton, Oregon 97006-6063 (503)526-4329 ...{tektronix,ogcvax}!sequent!sweiger -- Mark Sweiger Sequent Computer Systems Database Software Engineer 15450 SW Koll Parkway Beaverton, Oregon 97006-6063 (503)526-4329 ...{tektronix,ogcvax}!sequent!sweiger