Damasc can also be used to “emulate” (and to some extent, improve)
services provided by external libraries. We consider the recently
proposed PLFS [5] framework as an example. PLFS aims to
provide an efficient checkpointing mechanism for distributed applications.
It provides the abstraction of a single file for storing checkpoint
data, but internally associates each writer process of the distributed
application with a disjoint region of the file. Thus, concurrent
write requests can proceed concurrently without the need for
expensive synchronization. Our proposed file system can achieve a
similar functionality by using views. More concretely, each writer process updates a separate file (without the need for synchronization),
and the disjoint files are then union-ed in a view which presents
the abstraction of a single file. Recall that the view is not materialized,
and any access to the view is translated to accesses over
the constituent files. Moreover, the file system can automatically
index the contents of the view, thus improving the performance of
data retrieval. This is a definite advantage over PLFS, which cannot
support efficiently random-access reads over the single file.
Damasc can also be used to “emulate” (and to some extent, improve)services provided by external libraries. We consider the recentlyproposed PLFS [5] framework as an example. PLFS aims toprovide an efficient checkpointing mechanism for distributed applications.It provides the abstraction of a single file for storing checkpointdata, but internally associates each writer process of the distributedapplication with a disjoint region of the file. Thus, concurrentwrite requests can proceed concurrently without the need forexpensive synchronization. Our proposed file system can achieve asimilar functionality by using views. More concretely, each writer process updates a separate file (without the need for synchronization),and the disjoint files are then union-ed in a view which presentsthe abstraction of a single file. Recall that the view is not materialized,and any access to the view is translated to accesses overthe constituent files. Moreover, the file system can automaticallyindex the contents of the view, thus improving the performance ofdata retrieval. This is a definite advantage over PLFS, which cannotsupport efficiently random-access reads over the single file.
การแปล กรุณารอสักครู่..