|
|
|
|
|
|
|
Figure 6.1
The canonical computer system, consisting of multiple processors, cache
memories, and memory banks, connected by a processor-memory
interconnect. |
|
|
|
|
|
|
|
|
The Changing Role of Memory |
|
|
|
| |
|
|
|
|
In 1960, the cost for memory in a typical mainframe computer was about $1 per bit! Thirty years later, the cost has declined by a factor of a million, enabling memory sizes to increase by 110,000 and systems costs to decline by a factor of 100. |
|
|
|
| |
|
|
|
|
Access time and cycle time, however, have not changed nearly as dramaticallythey have decreased by perhaps a factor of 10. Cache memory has been a key development in the interim allowing improved processor performance. Large memory, in turn, has significantly ameliorated serious limitations in the I/O storage system. |
|
|
|
| |
|
|
|
|
Looking ahead, one would expect memory costs to continue to decline and sizes to increase, although not at the same rates that we have seen in the past. The principal limitation to current memory technology is volatility (the loss of information when power is removed from the memory chips). As memory sizes increase, the I/O storage system increasingly plays the singular role of a nonvolatile data storage backup system. |
|
|
|
|
|
|
|
|
|
In systems with multiple processors or with complex single processors, requests may cluster and congest the memory system. Either multiple requests may occur at the same time, providing bus or network congestion, or even within a single processor requests arising from different buffered sources may request access to the same memory module. Requests that cannot be immediately honored by the memory system result in memory systems contention. This contention degrades the bandwidth it is possible to achieve from the memory system. |
|
|
|
|
|
|
|
|
The most obvious factor in determining memory bandwidth is the number of memory modules into which the memory space is divided. The maximum theoretical bandwidth of the memory system is given by the number of memory modules divided by the memory cycle time. One might reasonably expect that increasing the number of memory modules would reduce |
|
|
|
|
|
