DISK CACHE ARCHITECTURES FOR TRANSACTION-LIKE APPLICATIONS IN PARALLEL COMPUTERS

I/O in computer systems is prone to become a bottleneck. This is a par ticular severe problem in highly parallel machines where some applicat ions are fully I/O bound if only one or few conventional I/O paths exi st. Similar to the use of multiprocessor technology for increasing pro cessing performance, disk I/O performance can be substantially improve d by employing parallel I/O schemes. Based on a distributed I/O archit ecture for parallel computers, we propose to use disk caches on severa l architectural levels, and confirm this by simulations of various str uctural options. In this paper, we describe the cache modelling approa ch and the I/O load model which has been derived From transaction-proc essing and general-purpose applications. Then we discuss the results f or caches on single and multiple architecture levels. Large caches on I/O processors in combination with small caches on processing elements turn out to be the preferable structure. In addition, hardware caches can be employed at disk level for further performance improvement. Fo r write operations, a delayed write strategy is shown to be superior t o other modes.