Matching Application Access Patterns to Storage Device ...

30 · Matching Application Access Patterns to Storage Device Characteristics
Dynamic decisions that depend on the current state of the storage device (e.g.,
current position of the read/write heads, or the content of prefetch buffers) should
be made below the interface. A storage manager should relegate these devicespecific
decisions (e.g., I/O request scheduling) to the storage device where they
can be made more efficiently. Instead, the SM is only concerned with generating
I/Os that can be executed efficiently. How these I/Os are executed, however, is
decided by the device.
The FFS storage manager example in Figure 3.2 includes a C-LOOK scheduling
algorithm (a variant of a SCAN algorithm [Denning 1967], which is implemented
as a part of the diskqsort() routine). The FFS scheduler makes its decisions on a
crude notion of relative distances in the LBN space. Instead, the scheduling should
be pushed down to the device, where it can be made more efficiently with more
detailed information (for instance, by using a SPTF scheduler [Seltzer et al. 1990;
Worthington et al. 1994], which is built into modern disk drive firmware [Quantum
Corporation 1999]).
3.1.4 LBN address space annotations
A storage device exports static performance attributes that annotate the linear
address space of fixed-size blocks, identified by a logical block number (LBN). This
annotation creates relations among the individual LBNs of the storage interface.
Given an attribute with particular semantics and a single LBN, other LBNs
satisfying the relation described by the attribute are returned.
This dissertation describes in detail two examples of these attributes and evaluates
the benefits these attributes have for a variety of applications and workloads.
– access delay boundaries
This attribute denotes preferred storage request access patterns (i.e., the
sets of contiguous LBNs that yield most efficient accesses). This attribute
encapsulates traxtent (an extent of LBNs mapped onto one disk drive track)
[Schindler et al. 2002] or a stripe unit in RAID configurations. This attribute
captures the following notions: (i) requests that are exactly aligned on the
reported boundaries and span all the blocks of a single unit are most efficient,
and (ii) requests smaller than the preferred groupings, should be aligned on
the boundary. If they are not aligned, they should not cross it.
– parallelism
Given an LBN, this attribute describes which other LBNs can be accessed
in parallel. The level of parallelism depends on the striping and RAID groups
and the number of spindles. For example, a logical volume of a storage array