Trail: Write-Optimized Disk Storage System
Group Members:
Trail is a disk storage system optimized for write performance.
The key idea of Trail is to take the concept of write anywhere to
the extreme so that the only cost of a physical disk write is the associated
data transfer delay. That is, there is neither queuing and seek delay, nor
rotational latency. Based on a Trail subsystem, the file system could be
optimized for read peformance, along the lines of BSD Fast file system.
Trail uses a log disk in addition to a normal disk. The log disk
speeds up the write performance, and together with the buffers at host memory,
forms a persistent cache at the device level. This cache services both reads and
writes, and is completely transparent to the file system's own buffer cache.
The key enabling technology of Trail is a low-overhead prediction
mechanism that can accurately estimate the disk head position on a given track.
Armed with this information, Trail is able to write to the log disk at
where the disk head happens to be, thus eliminating both seek delays and
rotational latency. This mechanism also serves as a crucial building block for
constructing rotation-sensitive disk schedulers, an increasingly important
technology as the rotational latency becomes the dominant component of physical
disk service time.
We have finished:
- Building a Linux-based Trail prototype.
- Integrate the Trail technology to a high-performance transaction
processing system to reduce the performance overhead of logging.
- Incorporate the Trail technology to build a rotation-sensitive disk
scheduler under Linux.
Publications:
- Tzi-cker Chiueh, "Trail: A Track-based Logging
Disk Architecture for Zero-Overhead Writes," In International Conference
on Computer Design, ICCD '93, October 1993, Boston, MA.
- Lan Huang, Tzi-cker Chiueh, "Charm: An I/O-Driven
High-Performance Transaction Processing System," USENIX Annual
Technical Conference(USENIX 2001), June 2001, Boston MA.
- Lan Huang, Tzi-cker Chiueh "Trail: Track-Based Logging in
Stony Brook Linux," in the Proceedings of the
2002 International Conference on
Dependable Systems and Networks (DSN 2002), June 2002, Washington, DC.
- Lan Huang, Tzi-cker Chiueh, "Experiences in Building a
Software-Based SATF Disk Scheduler," Technical Report
ECSL-TR81, revised in July, 2001.
Acknowledgement
This research is supported by an NSF Career Award
MIP-9502067, NSF MIP-9710622, NSF IRI-9711635, NSF EIA-9818342, NSF ANI-9814934,
a contract 95F138600000 from Community Management Staff's Massive Digital Data
System Program, USENIX student research gransts, as well as fundings from Sandia
National Laboratory, Reuters Information Technology Inc., and Computer
Associates/Cheyenne Inc.