Overall Write Cost

Write cost is the metric traditionally used in evaluating LFS write performance. It only considers the effect of the number of segments. Matthews et al. pointed out segment size also plays a larger role in the write performance. They described a way to quantify this trade-off between amortizing disk access time across larger transfer units and reducing cleaner overhead: Overall Write Cost, which captures both the overhead of cleaning as well as the bandwidth degradation caused by seek and rotational latency of log writes [9].

In this paper we used this new metric - overall write cost to evaluate WOLF performance. The following formula are adapted from [9]:

First, two terms, write cost and Transfer Inefficiency ($Ineff_{Xfer}$) are defined:

$$WriteCost = \frac{Segments~Transferred_{Total}}{Segments~Transferred_{NewData}}$$

$$=\frac{SegsW_{NewData}+SegsR_{Clean}+SegsW_{Clean}}{SegsW_{NewData}}$$

Here $SegsW_{NewData}$ is the total number of segments written to the disk caused by new data. $SegsR_{Clean}$ and $SegsW_{Clean}$ are the total numbers of segments read and written by the cleaner, respectively. This term describes the overhead of cleaning process.

$Ineff_{Xfer}$ measures the bandwidth degradation caused by seek and rotational delays of log writes. AccessTime represents the average disk access time.

And finally,

$${Overall~Write~Cost} = WriteCost \times {Ineff_{Xfer}}$$