The KK11-A is a cache memory option for the PDP-11/34A's KD11-EA processor.
The cache is a small, high speed memory that maintains a copy of previously selected portions of main memory (MOS or core).
The cache memory is designed to decrease central processing unit to memory read access time.

I read "somewhere" that the KK11-A was not such a big score, because *all* instructions get lengthened by several tens of microseconds. Only if you have a hit you gain performance, well that's usual the case with cache :-) , but "bad" programs (that experience lots of misses) will actually perform worse when the KK11-A is installed, because of the extra delay on all instructions.

Lyle Bickley has the definitive answer, he says the following.

Yes, I heard that rumor too. But when I studied DEC's own research on cache for PDP-11's - I concluded that the rumors were false.
It is kinda sad - but I held off implementing cache for a couple of years because I believed the rumors - as opposed to researching and finally benchmarking cache versus non-cache. You can read all about it in "Computer Engineering", Bell, Mudge and McNamara, © 1978, in Chapter 10 "Cache Memories for PDP-11 Family Computers".

Architecture

Cache operates as an associative memory in parallel with the UNIBUS main memory but with its own high-speed data path (AMUX lines that are also used by the FP11-A floating point processor option). Cache reads by the CPU result in data being transmitted over the AMUX lines.
Read misses (desired data is not present in the cache) and write hits (bus address and cache location match) which result in cache updates are accomplished by the cache capturing the data from the UNIBUS as the CPU/main memory transaction occurs. Direct memory access (DMA) transfers to memory are also monitored by the cache.

Performance
I received the following information from Lyle Bickley.
Lyle has done extensive benchmarks with and without the cache option installed in the 11/34 and he found that cache can make significant performance enhancements. Lyle says that he was surprised that cache helped performance as much as it did in his PDP-11/34 considering that his machine uses fast MOS (MS11-L) memory - and the cache was designed as a "kicker" for much slower memory.
Here is a brief summary of his findings of cache speed enhancement.

• Large FORTRAN compiles and MACRO assemblies run on average 44% faster.
• Large LINKs run 56% faster.
• Integer benchmarks like HANOI run 57% faster.
• Floating point benchmarks like single precision Whetstone run 19.4% faster and double precision Whetstone runs 16.9% faster.