BCW is based on the finding that HDDs are usually underutilized in an SSD-HDD hybrid storage server while SSDs are suffering from high write pressure.
The SSD delivers outstanding peak performance:
However, the long tail latency is also unneglectable:
The queue length of an SSD can be as large as more than 100.
HDD can reach 𝝁𝒔-level write latency, especially for small size requests:
Due to the HDD buffers, the HDDs repeatly performs fast/mid/slow writes:
The pattern looks like this: FFFF MMMS MMMS MMMS …
The latency of HDD buffered writes can be lower than SSD write latency and is predictable! Thus, the solution is to redirect writes to HDDs when profitable. This paper compared multiple redirect strategies.