The Free Lunch Has a Queue: Characterizing On-chip Compression Accelerators for Analytics
It is found that hardware offload is not a universal replacement for CPU compression, and compression should be scheduled dynamically: route blocks by codec, operation, size, and accelerator load; cap per-device sub-mission concurrency; and fall back to software when offload is unsupported or saturated.
Yi Jiang, Antonio Boffa, Hamish Nicholson et al.
· 0 citations