Database Reference
In-Depth Information
cache writes on clients, then dirty data might be lost, or operations might
have to be replayed from a log when the client returns.
Network failures can cause two problems. The most obvious is that they
can cause a single server (or client) to become inaccessible. This situation
is typically handled as if the server (or client) has died. A more complicated
situation arises when a network failure partitions the network, causing one set
of servers and/or clients to be unable to contact another set. This situation
is dangerous because if both groups continue to operate, the state in the two
groups could become out of synchronization. In order to avoid this situation,
file systems often use quorum techniques to prevent small groups of partitioned
servers or clients from continuing to operate. The quorum approach involves
passing messages routinely to establish that a given server can communicate
with at least half of the other servers (putting it in a majority). If it cannot,
then it shuts down, ensuring that it does not cause file system errors.
2.2 Parallel File Systems
While parallel file systems vary widely in the specifics of their implementation,
at a high level these systems can be split into two groups based on how clients
access data. In the first group, block-based storage systems, clients perform
I/O in terms of (real or virtual) disk blocks in a manner similar to a locally
attached disk drive. In the second group, object-based storage systems, clients
operate on regions of more abstract “objects,” containers holding data and
associated attributes. In this section we examine examples of both approaches.
2.2.1 Block-Based Storage Systems
Prior to the 1980s a popular data organization on disk was count-key-data
formatting, where applications decided how much data to put into each record
and what searchable key is associated with each record. In the early 1980s,
SCSI (small computer system interface) created a standard that kick-started
competition, leading to a much larger market with much better pricing; re-
sponded to a need for smaller form-factor disks for mini-, micro-, and personal
computers; and cost-effectively moved enough independent controller function
to the device to allow large collections to operate in parallel. In particular,
SCSI popularized the now common fixed-block disk format, in which every
disk record is the same size and is addressed by its index number in a lin-
ear space of N blocks—a memory abstraction. With this simple abstraction,
a linear block address space concealing the actual location of each block on
disk, came the opportunity to change the physical organization of blocks on
a disk, for example, to skip over a damaged block and use a spare block at
some other place whenever the damaged block's index is requested.
Search WWH ::




Custom Search