DAFS, Conquest May Pave Way for Future File Systems
Harvard student Kostas Magoutis presentation covered the functionality and benefits of DAFS (Direct Access File System), a powerful user-level file system backed by several vendors, including Network Appliance.
In a nutshell, DAFS makes storage networking faster by using a client side I/O library, which communicates directly with a servers NIC and avoids the kernel.
In benchmark tests, Magoutis found that in certain situations DAFS was able to reduce CPU overhead and increase transaction throughput when compared to NFS.
More information on this project can be found at www.eecs.harvard.edu/vino/fs-perf/dafs/.
One potential trouble spot is security. Although DAFS can probably work well in LAN-type settings, the scale of current networks and the fact that data typically runs though public and private networks to reach its destination requires that DAFS must work together with IPSec to ensure security.
The price of being different is that new technologies that are designed to enhance conventional technologies (like offloaded encryption hardware) may not work well with DAFS in the future, or may require a lot of tweaking to work.
Andy Wang, a student from UCLA, presented his work with Conquest, a hybrid file system utilizing disks and persistent (battery-backed) RAM.
Wang believes that the abundance of cheap, available RAM makes it possible to achieve high performance gains by storing metadata, small files, executables and shared libraries entirely on persistent RAM.
In Conquest, disk drives are only used to store large files (which in this design are classified as 1MB and bigger). Using Conquest, Wang ran benchmarks that showed improvements ranging from 24 percent to 1900 percent, depending on the type of workload and the amount of disk hits generated.
Wangs audience brought up a potential problem with the Conquest approach: There are many cases where the aggregate size of small files within a system is far larger than the standard RAM capacity of that system. A laptop with 5GB worth of small files, for example, would not have nearly enough RAM to hold all of these small files.
Another more important problem is the possibility of data getting corrupted while stored within RAM. If the stored data were corrupted, the system would not be able to recover from disk because metadata and small files only reside in memory.
More information on Conquest can be found at lasr.cs.ucla.edu/conquest/.
Although its too early to predict the impact these proposed technologies will have on the world of file systems, evolution is clearly going to continue in this field one way or another.
Senior Analyst Henry Baltazar can be reached at firstname.lastname@example.org.