Hidden productivity losses in computer systems
- PDF / 471,150 Bytes
- 10 Pages / 595 x 842 pts (A4) Page_size
- 26 Downloads / 235 Views
HIDDEN PRODUCTIVITY LOSSES IN COMPUTER SYSTEMS
UDC 681.324
Yu. S. Yakovlev
Based on an analysis of procedures of access to memory cells of computers and distributed computer systems (CSs), an estimate is given for time expenditures for address formation that do not belong to the category of direct expenditures for information processing in a processor and, hence, fall into the category of expenditures determining “hidden productivity losses” in a CS. Recommendations on the elimination of such losses during the creation of a CS are given. Keywords: computer system, productivity losses, addressing modes, address translation, memory access time. INTRODUCTION Over the entire period of development of computer facilities, numerous attempts were made in constructing supercomputers with a view to increasing the productivity of such a system not only by increasing the operating speed of its processors and their number but also by increasing their main memory speed and information transfer between a processor and its memory since an insufficient memory speed always slowed down the processor. New types of memory structures are created that provide accelerated access to information (for example, parallel multichannel access), new memory operation modes are proposed that maximally use the parameters of the timetable of processor operation (for example, read/write using “leading/falling edges” of clock signals), multilevel structures of cache memory (up to three levels) are developed, new technologies of creation of microelectronic components on a crystal and architecturally-structural solutions of a system as a whole are implemented that essentially extend the memory-processor bandwidth (for example, “Processor-in-Memory” architectures or PIM-systems), and memory networks with various architectural and structural organizations are used [1]. However, all technical and engineering solutions in this direction were considered without close interrelation with the system software whose memory control functions actually underwent minor changes from the moment of origin of operational systems [2]. As before, there are methods of paged, segmented, and segmented paged memory allocation and addressing, the problem of internal and external memory fragmentation is not solved, etc. This leads to the inefficient use of processors and losses in system productivity but such losses are “hidden” for structure developers and programmers since they are determined by distinctive features of functioning of an operational system. Investigations showed [3] that approximately 20–30% of the central processor time during the execution of applications are spent for allocation and cleaning of information in memory and approximately two thirds of the memory control time are spent for allocation. Memory fragmentation is an indication of memory use and actually determines the inability of the control program of memory allocation to use free space. Approximately 25% of the memory area is not used because of its internal fragmentation when memory is allocated
Data Loading...
