Chapter 4. Tuning the operating system 117
Draft Document for Review May 4, 2007 11:35 am 4285ch04.fm
cases the performance of the anticipatory elevator usually has the least throughput and
the highest latency. The three other schedulers perform equally good up to a I/O size of
roughly 16kB at where the CFQ and the NOOP elevator begin to outperfom the deadline
elevator (unless disk access is very seek intense) as can be seen in Figure 4-7.
Figure 4-7 Random read performance per I/O elevator (synchronous)
Complex disk subsystems
Benchmarks have shown that the NOOP elevator is an interesting alternative in high-end
server environments. When using very complex configurations of IBM ServeRAID or
TotalStorage® DS class disk subsystems, the lack of ordering capability of the NOOP
elevator becomes its strength. Enterprise class disk subsystems may contain multiple
SCSI or FibreChannel disks that each have individual disk heads and data striped across
the disks. It becomes be very difficult for an I/O elevator to anticipate the I/O
characteristics of such complex subsystems correctly, so you might often observe at least
equal performance at less overhead when using the NOOP I/O elevator. Most large scale
benchmarks that use hundreds of disks most likely use the NOOP elevator.
Database systems
Due to the seek-oriented nature of most database workloads some performance gain can
be achieved when selecting the deadline elevator for these workloads.
Virtual machines
Virtual machines, regardless of whether in VMware or VM for System z, usually
communicate through a virtualization layer with the underlying hardware. Hence a virtual
machine is not aware of the fact if the assigned disk device consists of a single SCSI
device or an array of FibreChannel disks on a TotalStorage DS8000™. The virtualization
layer takes care of necessary I/O reordering and the communication with the physical
block devices.
CPU bound applications
While some I/O schedulers may offer superior throughput than others they may at the
same time also create more system overhead. The overhead that for instance the CFQ or
deadline elevator cause comes from aggressively merging and reordering the I/O queue.
0
20000
40000
60000
80000
100000
120000
140000
160000
180000
kB/sec
4 8 16 32 64 128 256 512 1024 2048
kB/op
Deadline
Anticipatory
CFQ
NOOP
