💨

[Cache Compression] The XOR Cache

に公開

Key Contribution

reducing requirement of LLC(Last Level Cache) reduce accessing RAM by Cache miss

XOR Compression

compress B by XOR
Delta have many zeros

Delta = A XOR B
  0011  (A)
^ 1011  (B)
----------
  1000  (Delta)

Reconstruct B by XOR

B = A XOR Delta
  0011  (A)
^ 1000  (Delta)
----------
  1011  (B)

Inter-Line Compression

by keeping Delta, A and B can be Cache Coherent. which makes 1/2 cahce line size in LLC(Last Level Cache)

Intra-Line Compression

Delta can be compressed by Sparse Algorithms such as Zero-Value Compression which keep indices of 1s

ex)

1000000000000001 -> (15,0)

Reference

The XOR Cache: A Catalyst for Compression

Discussion