热点感知分组弹性条带化提升SSD RAID性能
81 浏览量
更新于2024-08-28
1
收藏 526KB PDF 举报
"基于热点感知的分组弹性条带化，以提升SSD RAID性能的研究论文"
本文探讨了如何通过一种新的方法来优化固态硬盘（SSD）RAID（冗余磁盘阵列）的性能，该方法被称为“基于热点感知的分组弹性条带化”。传统的RAID在提供设备级容错功能时，通常采用读-修改-写或读-重建-写的方式更新奇偶校验块，这种操作会导致大量额外的I/O操作，从而显著降低SSD RAID的性能。
最近提出的弹性条带化方案尝试用更新的新数据块重建新条带，而不更新旧的奇偶校验块，以此减少不必要的I/O操作。然而，这种方法仍然需要在RAID级别进行垃圾回收，这可能会带来很高的成本。为解决这个问题，作者提出了一个基于热点感知的缓存策略，用于缓冲即将到来的写入操作，并根据数据块的热度值将它们分类到多个缓冲区组中。
“热点”是指在系统中访问频率较高的数据，对这些高访问频率的数据进行特殊处理可以提高整体性能。在这个热点感知的缓存策略中，不同热度级别的数据块被分组到不同的缓冲区，这样可以更有效地管理和调度I/O操作。随后，他们提出了一种分组式的弹性条带化方案，将不同组内的数据块分别写入到不同的存储区域，以此避免因频繁的垃圾回收操作导致的性能下降。
这种分组弹性条带化策略的核心思想是根据数据的热度和访问模式优化写入操作，使得热门数据能够在写入时优先考虑，从而减少了I/O延迟和提高了系统的整体响应速度。同时，通过合理地组织和更新数据，可以降低对RAID级别的垃圾收集需求，进一步提升SSD RAID的性能。
此外，文章可能还深入分析了该方法的实现细节，包括如何评估和更新数据块的热度值、如何动态调整缓冲区大小以及如何在不同组间平衡负载等。可能还涉及了性能评估，通过实验对比传统的RAID技术和现有的弹性条带化技术，证明了所提出的策略在实际应用中的优越性。
这篇研究论文提供了一个创新的解决方案，旨在利用热点数据的特性提高SSD RAID的性能，通过智能缓存管理和分组弹性条带化策略，有效减少了不必要的I/O操作，降低了系统开销，提升了固态硬盘阵列的效率。
SSD 0
Valid chunkInvalid chunk
SSD 1 SSD 2 SSD 3
0
D
1
D
2
D
0
P
SSD 0 SSD 1 SSD 2 SSD 3
0
D
1
D
2
D
0
P
SSD 0 SSD 1 SSD 2 SSD 3
0
D
1
D
2
D
0
P
1
D
2
D
0
P
1
D
2
D
0
P
1
D
2
D
(a) (b) (c)
1
D
2
D
3
D
4
D
5
D
1
P
4
D
1
P
3
D
4
D
5
D
1
P
3
D
4
D
5
D
1
P
4
D
1
P
4
D
2
P
4
D
Log
device
1
L
2
L
4
L
SSD 0 SSD 1 SSD 2 SSD 3
0
D
1
D
2
D
0
P
SSD 0 SSD 1 SSD 2 SSD 3
0
D
1
D
2
D
0
P
3
D
4
D
5
D
1
P
3
D
4
D
5
D
1
P
Log
device
Incoming requests: 1. Update 2. Update 3. Update
Before
After
1
*
D
SSD 0 SSD 1 SSD 2 SSD 3
0
D
1
D
2
D
0
P
3
D
4
D
5
D
1
P
Cache
Cache
2
*
D
4
*
D
Note that: Note that: Note that:
1 1 1
L D D !
2 2 2
L
D D !
4 4 4
L
D D !
0 0 1 2
P P L L ! !
1 1 4
P P L !
0 0 1 2
P D D D ! !
1 3 5 4
P D D D ! !
2 1 2 4
P D D D ! !
Fig. 2: Three categories of enhanced RAID schemes: (a) Parity Logging; (b) Parity Caching; (c) Elastic Striping
a round-robin manner for load balance. When data chunks are
updated, their corresponding parity chunks will be updated as
well via either read-modify-write or read-reconstruction-write.
We let the chunk size be equal to the page size in this paper.
C. Enhanced RAID Schemes
As we discussed in
§
I, parity update in a RAID will
introduce extra I/Os and degrade both the performance and en-
durance of SSD RAID. Various RAID schemes are developed
to reduce the I/Os caused by parity update, and we classify
them into three categories: parity logging [29, 23, 36], parity
caching [7, 13, 18, 20], and elastic striping [16].
Fig. 2 illustrates the three schemes. Suppose that there are
six data chunks D
0
, D
1
, D
2
, D
3
, D
4
, D
5
and two parity chunks
P
0
, P
1
stored in an SSD RAID array at the beginning, and the
incoming requests are: (1) updating D
1
to D
∗
1
, (2) updating
D
2
to D
∗
2
, (3) updating D
4
to D
∗
4
. We assume that the three
update requests arrive sequentially.
Fig. 2(a) illustrates the parity logging scheme. It usually
employs a dedicated device to log data writes, e.g., uses an
HDD as the log device to absorb small writes. With parity
logging, D
1
is updated to D
∗
1
in SSD 1, and a delta, L
1
,
which is computed via XOR-ing D
1
with D
∗
1
, is logged on
the log device. Then D
2
is updated to D
∗
2
in SSD 2, and
L
2
is also logged. When serving the last request, it writes
D
∗
4
into SSD 1, and writes L
4
into HDD. At last, it uses the
defined computation denoted in Fig. 2(a) to update P
0
to P
∗
0
and P
1
to P
∗
1
in SSD 3 and SSD 2, respectively. We note that
the number of parity writes gets reduced, so as the writes to
SSDs. However, parity logging requires a dedicated device to
log small writes, and it may also harms system-level wear-
leveling.
Fig. 2(b) shows an example of the parity caching scheme.
Parity caching uses a buffer to cache all incoming writes so to
delay parity updates. The key idea is that if more chunks are
updated together, the chance of constructing full-stripe writes
becomes larger, so the I/Os caused by parity update can be
reduced. As shown in Fig. 2(b), it buffers D
∗
1
and D
∗
2
in the
cache first, and then flushes them to SSDs and updates P
0
to P
∗
0
. This process will also happen when updating D
4
and
P
1
. Parity caching can also reduce parity writes, but it usually
requires a dedicated NVRAM, which is expensive and still not
mature in the current market. Note that D
1
, D
2
, D
4
, P
0
and P
1
in both Fig. 2(a) and Fig. 2(b) are marked invalid at the device
level after data update and they can be reclaimed automatically
by the corresponding SSDs.
As shown in Fig. 2(c), elastic striping manages write
requests in a log-structured manner. It appends D
∗
1
, D
∗
2
and D
∗
4
into the RAID array for constructing a new stripe instead of
updating them in the original devices. Note that D
1
, D
2
and D
4
are out-of-date but are still needed to be kept in SSDs for data
protection, and this consumes extra storage space. Therefore,
elastic striping marks these chunks as invalid at RAID level and
calls RAID-level GC operation to reclaim the space occupied
by these invalid chunks. This scheme can effectively reduce
parity writes, mitigate the performance degradation caused by
RMW and RRW, and benefit the wear-leveling among SSDs.
Moreover, elastic striping seems to be suitable for SSD RAID
as out-of-place overwrite is usually adopted by SSDs. The
primary issue of this scheme is that RAID-level GC cost may
be very high if invalid chunks are scattered over the stripes
in the whole RAID array. This motivates us to develop a
workload-aware scheme to reduce the RAID-level GC cost so
to improve the SSD RAID performance. We further discuss
162162
剩余11页未读，继续阅读
2022-05-02 上传
2022-07-13 上传
2021-02-09 上传
2021-06-30 上传
2017-05-03 上传
2022-06-25 上传
2022-07-12 上传
2022-07-12 上传
weixin_38663169
- 粉丝: 2
- 资源: 915
我的内容管理
展开
最新资源
- Aspose资源包：转PDF无水印学习工具
- Go语言控制台输入输出操作教程
- 红外遥控报警器原理及应用详解下载
- 控制卷筒纸侧面位置的先进装置技术解析
- 易语言加解密例程源码详解与实践
- SpringMVC客户管理系统：Hibernate与Bootstrap集成实践
- 深入理解JavaScript Set与WeakSet的使用
- 深入解析接收存储及发送装置的广播技术方法
- zyString模块1.0源码公开-易语言编程利器
- Android记分板UI设计：SimpleScoreboard的简洁与高效
- 量子网格列设置存储组件：开源解决方案
- 全面技术源码合集：CcVita Php Check v1.1
- 中军创易语言抢购软件：付款功能解析
- Python手动实现图像滤波教程
- MATLAB源代码实现基于DFT的量子传输分析
- 开源程序Hukoch.exe：简化食谱管理与导入功能
