Input-output strategy

User programs that needed access to amounts of data larger than the amount that could be swapped into central memory would use some form of access to data stored in ECS. (Eventually called ECS files, see next chapter.) This data stored in ECS could then be used as an I-O buffer. Programs which desired access to this data would read it while they were running in central memory. Data from an input-output device would first go to a PPU. The PPU would transfer the data to a central memory buffer and start a system program in central memory, which would transfer the data to ECS. This system program would then inform the appropriate requesting program that the data was available. (The requesting program was informed by sending an event on an event channel, see next chapter.) Movement in the opposite direction was similar. We made rough estimates of the amount of CPU time that would be used by the system programs to move the data to and from ECS. We assumed that the main I-O load would come from the disk. If we assumed that two PPU's were alternately reading the disk and then writing into central memory, central would receive words at the maximum rate of one word every five microseconds. If enough were buffered so that the main cost to the central program was the actual transfer to ECS, the central program would use one microsecond every 50. Thus, we felt this time overhead was acceptable. We gave less thought to the space overhead. However, at 512 words per PPU (the maximum power-of-2 buffer that a PPU could hold), there would be at most 5120 words of central. On a 64K machine (which we assumed would be available for such a large system) this is less than 10 the total space, which we thought would be acceptable. If this was too much, we could arrange for the PPU's to transfer a 512 word buffer in several sections, with a separate request to central memory system program for each one. This would reduce the necessary buffer space in central memory at an increased cost for CPU time, due to overhead in responding to each request. As an alternative to this design with central memory buffers, CDC offered an optional hardware feature which permitted direct exchange of data between a PPU and ECS [C2]. We referred to this feature as a ``Back-door''. Unfortunately, the rate of this transfer was limited, by the PPU, to one 60 bit word every 5 microseconds, and while in progress degraded the transfer rate between central memory and ECS. We were unable to determine from CDC how great the degradation would be, but there were hints it might be a factor of two. Since our estimates of program swapping overhead were heavily dependent on a high transfer rate between central memory and ECS, and the estimates for the PPU-CM-ECS scheme indicated a low overhead, we decided to avoid the Back door.