Proving Kangaroo Works Efficiently

Results

Let’s look at the results of Kangaroo’s design and optimization in practice.

Consider we have the following graph where we have miss ratio and days on the y-axis and x-axis, respectively. We are comparing miss ratio for all three systems; log-structured, set-associative cache, and Kangaroo’s; over a 7-day Facebook trace. All systems are running with 16 GB DRAM, a 1.9 TB drive, and with write rates less than 62.5 MB/s.

The worst performance out of these three options is the log-structured cache and this is because it’s severely constrained under our setup. Even with being generous in terms of DRAM to log-structured cache, it cannot index the whole flash capacity and thus the miss ratio is not very great.

With the set-associative cache, it’s not as bound by the DRAM capacity but it is write constrained so it can also not do well in terms of the miss ratio.

Kangaroo is able to span both the DRAM and the miss rate constraints. It allows us to have a 29 percent miss ratio reduction over the social cache and an even larger miss reduction over the log-structured cache. It can span both the write amplification and the DRAM constraint while also implementing a better eviction policy to allow it to have a better miss reduction.