Abstract—The read/write performance asymmetry of Solid State
Disks (SSDs) remains a critical concern for read performance.
Under a concurrent workload with a mixture of read and write
requests, preceding write requests preempt available flash memory
resource so as to block read requests, which we call write-caused
interference. Hence, the read performance, which is often more
critical than the write performance, can be significantly degraded.
Unfortunately, state-of-the-art schedulers either are inefficient in
improving the read performance or suppress the write performance.
In this paper, we propose a novel scheduler at device level, called
AOS, to mitigate the write-caused interference and maximize the
read performance without sacrificing the write performance.
Specifically, AOS designs a conflict detection module, to efficiently
identify access conflicts among requests. Then, AOS adaptively
dispatches as many outstanding requests as possible to a re-ordering
set based on the detected conflicts to reduce the write-caused
interference and improve the flash-level parallelism (FLP). Finally,
AOS carefully re-orders the dispatched requests to reduce
channel-level access conflicts and improve the system-level
parallelism (SLP). Extensive experimental results show that AOS
reduces, an average of 51% read latency and 45% write latency,
compared to FIFO.
Index Terms—out-of-order scheduling, solid state disk,
write-caused interference, conflict detector
I. I
NTRODUCTION
AND flash-based Solid State Disks (SSDs) have been
widely deployed in data centers due to higher
throughput, lower latency, and lower energy than hard
disk drivers (HDDs) [1]. However, SSDs suffer from two
critical limitations: the read/write performance asymmetry
and the erase-before-write feature. Previous works [2, 3]
Manuscript received December 11, 2014; revised January 20, 2015. This
work is sponsored in part by the National Natural Science Foundation of
China No. 61300047 and the National Basic Research Program of China
(973 Program) under Grant No.2011CB302303。
Pingguo Li is with the Wuhan National Laboratory for Optoelectronics,
Huazhong University of Science and Technology, Wuhan 430074, P.R.
China, and the Library, Hubei University of Science and Technology,
Xianning 437000, P.R.China ( e-mail: pingguoli@hust.edu.cn
)
Fei Wu is with the Wuhan National Laboratory for Optoelectronics,
Huazhong University of Science and Technology, Wuhan 430074, P.R.
China (phone: 0086-27-87792405; fax: 0086-27-87792405; e-mail:
wufei@hust.edu.cn
).
You Zhou is with the Wuhan National Laboratory for Optoelectronics,
Huazhong University of Science and Technology, Wuhan 430074, P.R.
China (e-mail: zhouyou@hust.edu.cn).
Changsheng Xie is with the Wuhan National Laboratory for
Optoelectronics, Huazhong University of Science and Technology, Wuhan
430074, P.R. China (e-mail: cs_xie@hust.edu.cn
)
Jiang Yu is with the IBM China System and Technology Group Lab,
Shanghai 200000,P.R.China (e-mail: jiangyu@cn.ibm.com)
show that, under a concurrent workload with a mixture of
read and write requests, the read performance can be
significantly degraded, since preceding writes requests
preempt available flash memory resource and block read
requests. Note that the write operation of flash memory is
much slower than the read operation. To make the matter
worse, if the block to be written is not free, an erase operation,
which is much slower than the write operation, must be
performed before serving the write request due to the
erase-before-write feature, further degrading the read
performance. We refer to the read performance degradation
as the write-caused interference. Unfortunately, the read
performance is often more critical than the write performance
[4], it is necessary to mitigate the write-caused interference.
In order to achieve this, some studies on SSD I/O schedule
are proposed at operating system level. Wang et al. [5]
attempted to divide incoming read/write requests into
different sub-regions, and then served them in a round-robin
manner. Gao et al. [6] dispatched requests to different batches
based on an access conflict detection approach to avoid
access conflicts. Although these works achieve some
optimizations in reducing the write-caused interference, the
optimizations are purely based on the logical page addresses
(LPAs) required by the file system. Since the same LPA is
always redirected to different physical page addresses (PPAs)
due to the out-of-place write strategy of SSDs [2, 3], the
optimizations cannot be efficient and accurate. To avoid these
drawbacks, Wu et al. [7] proposed a P/E (program/erase)
suspension, a device-level scheduler, to reduce the
write-caused interference. However, frequently
suspending/resuming the on-going P/E operation introduces
system overhead and suppresses the write performance. In
addition, this method required hardware modification.
In this paper, we propose AOS, a novel device-level
scheduler, to mitigate the write-caused interference and
maximize the read performance without sacrificing the write
performance. This paper makes the following contributions:
We proposed a conflict detection module to identify access
conflicts among requests.
We proposed an adaptive request dispatching policy to
dispatch as many outstanding requests as possible to a
re-ordering set to reduce the write-caused interference and
improve the flash-level parallelism (FLP). Our experiments
show that this policy reduces an average of 55%
write-caused interference and improves FLP by about 2
times over FIFO under enterprise workloads.
We proposed a re-ordering policy, which reorders the
dispatched requests in a round-robin manner to reduce
channel-level access conflicts and improve the system-level
parallelism (SLP). Our experiments show that this policy
reduces an average of 31% channel-level access conflicts and
AOS: Adaptive Out-of-order Scheduling for
Write-caused Interference Reduction in Solid
State Disks
Pingguo Li, Fei Wu*, You Zhou, Changsheng Xie, Jiang Yu
Proceedings of the International MultiConference of Engineers and Computer Scientists 2015 Vol I,
IMECS 2015, March 18 - 20, 2015, Hong Kong
ISBN: 978-988-19253-2-9
ISSN: 2078-0958 (Print); ISSN: 2078-0966 (Online)