Maynard Handley (name99.delete@this.name99.org) on August 19, 2018 11:12 am wrote:
> Paul A. Clayton (paaronclayton.delete@this.gmail.com) on August 18, 2018 6:19 pm wrote:
[snip]
> > All zero cache lines (and pages) compress extremely well. I think it
> > makes more sense to use metadata to indicate such compression (at page
> > granularity one could steal a state in PTEs, which further avoids excess
> > cost in metadata fetching) and avoid memory read entirely. The lines can
> > be zero filled around the first write.
> >
>
> What you say is true, of course, but it's orthogonal to the aspect I wanted to highlight, the use of "remote
> atomics" for more than just atomic operations.

Remote (and parallel) operations can apply to other functions.
Memory management (particularly freeing memory, which has been proposed
as a means of modestly increasing thread-level parallelism) is a somewhat
obvious function. Many "background tasks" could be offloaded from less
energy-efficient ILP processors, sometimes exploiting data locality.
(Moving computation to data is another general concept, including
processing-in/near-memory and task assignment. Network topology seems
likely to offer similar though smaller optimization opportunities compared
with data locality.)

> It fits more into the "compressed L3" suggestion.

I agree.

> Zeroing is simply an especially simple example, but the other real world examples (memfill,
> and then memcopy) build on top of it, and are usually less amenable to simple compression.
> (memfill maybe, depending on your line compressor; memcopy likely not).

Copy engines (with fill capability) are a well-known off-loading, though
they also seem less used than possible.

Memcopy does present the opportunity for dictionary compression as there will
(temporarily) be at least two copies of the data. At least one NUCA
proposal used pointers to cache blocks in tags; aligned (or half-aligned
using overlaid skewed associativity to have different ways map to
different alignments) cache lines could be trivially copied with such
pointers.

(Cache compression might also interact positively with versioned memory
where the content between different versions may have significant
similarity.)

Memory copy has more than one mechanism for optimization, depending on
size, current location, and next use location and timing.

This also points to how optimizations tend to draw various aspects
together. This is partially to be expected as reducing abstraction
penalties is a common optimization perspective. However, exposing
interactions makes the system more complex (increasing design
effort, performance variation [increasing the importance of
software optimization and the chance for timing-based information
leakage], and risk of performance and correctness bugs).