Log-Performance Modeling

Bark is a layer that sits between the file system and disk and redirects journal writes so as to reduce rotational overhead. To do so, Bark builds a performance model of the log a priori and uses it to decide where best to write the next log write.

Our approach builds on our previous work that measures the request time between all possible pairs of disk addresses in order to perform disk scheduling [23]. Our problem here is simpler: Bark must simply predict where to place the next write in order to reduce rotation.

To make this prediction, Bark performs measurements of the cost of writes to the portion of the disk of interest, varying both the distance between writes (the ``skip'' size) and think time between requests. The data is stored in a table and made available to Bark at runtime.

For the results reported in this paper, we created a disk profile by keeping a fixed write size of 4 KB (the size of a block), and varying the think time from 0 ms to 80 ms in intervals of 50 microseconds, and the skip size from 0 KB to 600 KB in intervals of 512 bytes. To gain confidence each experiment was repeated multiple times and the average of the write times was taken.

Table: Bark Performance. Each row of the table plots the overall performance (in seconds) of TPC-B in three different settings: without Bark, with Bark, and on a ``null'' journal that reports success for writes without performing disk I/O (the null journal represents the best possible improvement possible by using Bark). The first row reports performance of a cold run, where table reads go to disk. The second row reports performance when the table is in cache (i.e., only writes go to disk). Experiments were run upon a Sun Ultra20 with 1 GB of memory and two Hitachi Deskstar 7K80 drives. The average of three runs is reported; there was little deviation in the results.

	w/o Bark	w/ Bark	Null
Uncached	50.7	42.1	38.8
Cached	44.2	27.3	25.4

0.9