How many additions per second (or, generally, instructions per second) could the IBM System370 Model 115 perform?

Question

I can find a lot of information about that model's memory access time, but not so much about its actual computational speed. I'm interested in the number of additions per second if possible, but I'll accept clock speed or something similar if that's all I can get.

Answer 1

I can find a lot of information about that model's memory access time, but not so much about its actual computational speed.

It's a commercial mainframe, an EDP system, not a number cruncher. in EDP it's all about data shovelling:

1. Read record
2. multiply item number by item price
3. print record.
4. repeat 1..3 until EOF

I'm interested in the number of additions per second if possible

Well, what additions?

Integer or Float? Long or short? Register/Register or Register/Memory?

:))

Also, while not a top end model, the -115 did already benefit from clever micro programming, cache, virtual addressing and interleaved memory, so either of that failing (not a cache hit, memory tables not in TLB, etc.) would increase execution time greatly.

According to a table in its 'Hardware Manual' (IBM System/370 Model 115 Functional Characteristics p.199 ff) most basic instructions range between 4 and 20 microseconds. A basic register register add (AR R1,R2) for example would be performed in about 8 µs giving a theoretically speed of 125 kOPs. Doing the same with a memory operand would be at least 14 µs plus optional indexing. A floating point addition would be anywhere between 40 and 65 µs, depending on operation type and operand location.

Of course, this differed as well over the lifetime of the -115. For example later models (like -F* and -G*) did speed up execution by 15%. Here an AR was done in a bit over 7 µs (see the 1976 document p.250 ff)

So go ahead and pick any model and instruction you like to investigate.

For all practical use these numbers were rather irrelevant. EDP is, as mentioned, about data shovelling. In fact, even memory speed itself is only half the story, as it's even more about I/O speed. The -115 could do a sustained I/O transfer of about 2 MiB per second. This is what defines the performance of an EDP system - Processes are most of the time waiting for I/O.

For all practical purpose a -115 can be rated at about 10-20 kOPs.

Which is quite fine for a very low end machine - it was the entry model of the /370 range, the smallest offering virtual memory. In fact, the -115 wasn't even present when the /370 was introduced in 1970, but came three years later (1973), even after the already downsized -125 (1972). It's over all speed was about 1/4th of a -135 mainly reached by cutting the memory interface from 8 byte to 2 byte. Due this lower data rate, the CPU timing was reduced to about 1/3rd on average.

It's main target was to convert the lowest end of /360 users, owners of /360-20, over to the new virtual addressing world and default use of disk drives. The -20 was was not even a full /360 CPU. It had only half the register set and a subset of instructions. It's main purpose was to replace mechanical punch card based systems.

Remember: Read Card - Add Price - Print line.

While by numbers the most successful /360, support was PITA for IBM as all software had to take be tailored to this machine, adding serious constrains to IBM's software and OS development.