没有合适的资源?快使用搜索试试~ 我知道了~
Optimizationsfor High Performance Network Virtualization
0 下载量 61 浏览量
2021-02-10
14:54:16
上传
评论
收藏 200KB PDF 举报
温馨提示
The increasing requirements of intensive interoperaterbility among the distributed nodes desiderate the<br>high performance network connections, owing to the substantial growth of cloud computing and datacenters.<br>Network I/O virtualization aggregates the network resource and separates it into manageable parts for particu<br>lar servers or devices, which provides effective consolidation and elastic management with high agility, flexibility<br>and scalability as well as reduced cost and c
资源推荐
资源详情
资源评论
. JOURNAL OF COMPUTER SCIENCE AND TECHNOLOGY : 1– Mon. Year
Optimizations for High Performance Network Virtualization
Fanfu Zhou, Ruhui Ma, Jian Li, Lixia Chen, Weidong Qiu, Haibing Guan
∗
Shanghai Key Laboratory of Scalable Computing and Systems,
Shanghai Jiao Tong University, Shanghai, 200240, China
E-mail: {zhoufanfu, ruhuima, li-jian, trollyxia, qiuwd, hbguan}@sjtu.edu.cn
Abstract
The increasing requirements of intensive interoperaterbility amon g the distributed nodes desiderate the
high performan c e network connections, owing to the substantial growth of cloud computing and datacenters.
Networ k I/O virtualization aggregates the network resource and separa tes it into manageable parts for particu-
lar servers or devices, which provides effective co n so lidation an d elastic management with high agility, flexibility
and scalability as well as reduced cost and cabling. However, both network I/O virtualization aggregation and
the increas in g network speed incur higher traffic density, that gener at es a heavy system stress for I/O data
moving and I/O event processing. Consequently, many researchers have dedicated to enhancing the system
performance and alleviating the system overhead for high performance networking virtualization. This paper
first elaborates the mainstreaming I/O virtualization methodolo g ies, including device emulation, split-dri ver
model and hardware assisted model. Then, the paper discusses and compares their specific advantages in addi-
tion to perf or ma n c e bottlenecks in practical utilities. This paper mainly focuses on the comprehensive survey
of state-of-the-art approaches for performance optimizations and improvements as well as porta b ility manage-
ment for network I/O virtualization. The approaches include var io u s novel data delivery schemes, overhead
mitigations for interrupt processing and adequate resource allocation s for dynamic network states. Finally,
we highlight diversity of I/O virtualization besides the performance improvements in network virtualization
infrastructure.
Keywords Networ k Virtualizat io n , SR- I OV, VMM
1 Introduction
The past decade has witnessed the rapid de-
velopment of cloud computing and datacenter
infrastructures, meanwhile, more and more in-
tensive data exchanges are required among th e
distributed servers in order to accomplish com-
plex event proces s in g [1], business-related data
∗ Corresponding Author
2 J. Comput. Sci. & Technol., Mon.. Year, ,
processing [2], decision support, video conferenc-
ing, map-reduce and graph compu tin g tasks [3],
etc. According to a Cisco report, the amount of
global datacenter IP traffics will nearly quadrupl e
over the next five years, growing at a compound-
ed annual growth rate of 31% [4]. High inten-
sive collaborative workloads require higher physi-
cal bandwidth consumption, so high performance
network devices are widely adopted in datacen-
ter networks. Currently, 10Gigabit per second
(Gbps) Ethernet network interface car ds (NICs)
have become the mainstream network devices in
the cloud computing and datacenters [5, 6] and
40Gbps-100Gbps NICs are in the development or
close to deployment [7, 8].
Input/Output (I/O) virtualization abstracts
upper layer protocols from physical connections
and dynamically shares the high-speed intercon-
nect among various requirements without any re-
strictions on physical port counts [9, 10]. I/O vir-
tualization consolidates many connection traffics
to a single physical link to achieve high network
resource utilization, and simpli fie s the manage-
ment of the I/O devices with software-based con-
figuration [11, 12]. These densified tr affics can ef-
fectively use fewer cards, cables, and switch port -
s, so as to s igni fic antly reduce datacenter server-
to-network and server-to-storage cabling by more
than 70 percent [13], and significantly lower the
cost, complexity, and power requirements [12],
accordingly. Therefore, I/O virtualization tech-
nology achieves outstand ing behaviors in terms of
greater flexibility, greater utilization and faster
provisioning comparing to the traditional NIC
and HBA (host bus adapter) card architectures
[14].
The continuously increasing speed and den-
sity of network traffics result in a higher amoun-
t of data movement and higher frequency of
network event n otifi cation s through the networ k
transmission path. Moreover, the multiplexing
of virtual machines (VMs) on sharing physical
server resource causes frequent context switch.
This costly overhead inflicts the network I/O ef-
ficiencies and responsiveness seriously. There-
fore, effectiveness and efficiency of high perfor-
mance network virtualization are critical for a
virtualization-based cloud computing and dat-
acenter systems [15]. To this end, n ovel net-
work data transmission schemes, effective net-
work e vent notification processing mechanisms
and dynamic system scheduling methods are high
desired to improve the network I/O virtualiza-
tion performance, especially for 10Gbps Ethernet
NICs and higher sp ee d NICs connections [16].
Generally sp eaki ng, I/O virtualization can
be classified into three methods according to the
network virtualization proce ss , such as device
emulation model, split driver model, and hard-
ware assisted model. In the full virtualization
scheme, the hyper v is or emulates device hardware
with clean abstraction, but the I/O operation
of guest OS needs costly and complicated trap-
and-emulation process to execute the lowest level
of the conversation [17]. The para-virtualization
(PV) scheme employs s p lit driver model to set
In this article, the term guest is interchangeable w ith guest domain, domU and guest OS, hypervisor is interchange-
able with VMM (virtual machine monitor).
Shortened title within 45 charactors 3
up a more efficient cooperation channels betwee n
the guest OS and the hypervisor. However, the
guest in PV needs to be modified to be aware of
its virtualized status, while the guest OS in ful-
l virtualization is unaware of virtualization and
requires no changes to work [18]. More recent-
ly, hardware assisted I/O virtualization offloads
the device emulation functions to specific hard-
ware, and provides natively sh ar eabl e devices for
each VM with direct access to I/O device without
VMM involvement, such as single- and multi-root
I/O virtu aliz ation proposed by PCI-SIG [19].
In this survey, we comprehensively and pro-
foundly discuss the challenges for high perfor-
mance network virtualization and the require-
ments according to specific high performance
network virtualization architectures. Also, we
make a detailed survey of the state-of-the-art ap-
proaches for performance optimizations based on
data moving, interrupt handling and resource al-
location. Besides, we state the management en-
hancement as well as diversity of I/O virtualiza-
tion.
The remainder of the paper is organized as
follows. Section 2 introduces the bas ic I/O vir-
tualization model s and their different I/O prop-
erties. Afterwards, we conclude th e performance
bottlenecks and challenges in Section 3. Section
4 concludes and summarizes the s t ate-of- th e-ar t
research strategies about performance optimiza-
tions and improvements for n etwork I/O vir t ual-
ization, as well as discusses the diversity of I/O
virtualization. Finally, we conclude in Section 5.
2 Network I/O Virtualization Models
Virtualization technology has a long devel-
opment history for logically dividing the system
resources provided by mainframe computers b e -
tween different applications [20, 21]. I/O vir-
tualization is a critical component to achieve the
successful and effective v ir t uali zed servers, and it
is recognized that 75% of virtualized servers re-
quire seven or more I/O connections per device,
and require more frequent I/O reconfigurations
[22]. In what follows, we will introduce th r e e I /O
virtualization models: device emulation model, s-
plit driver model and hardware assisted model.
2.1 Device Emulation
Device emulation method for I/O virtualiza-
tion is normally employed in full-virtu aliz ation,
which mimics I/O oper ati ons of the physical de-
vices by virtualizing sens iti ve and privileged in-
structions in VMM layer. This effect fully ab-
stracts and decouples the underlying hardware
from guest OS with the combination of bina-
ry translation and direct execution techniques.
Device emulation method offers the best per-
formance isolation and security for I/O opera-
tions, and simp lifi e s migration and portability
as guest OS instance and user level instructions
run u nmodified without awareness of virtualiza-
tion. The device emulation method using binary
translation has become the de facto standard in
VMware produc t series, such as VMware Work-
station [23] and VMware ESX Server [24]. In
addition, QEM U [25] of Xen [18] and KVM [26]
can also be configured to deploy device emulation
剩余13页未读,继续阅读
资源评论
weixin_38663036
- 粉丝: 4
- 资源: 928
上传资源 快速赚钱
- 我的内容管理 展开
- 我的资源 快来上传第一个资源
- 我的收益 登录查看自己的收益
- 我的积分 登录查看自己的积分
- 我的C币 登录后查看C币余额
- 我的收藏
- 我的下载
- 下载帮助
最新资源
资源上传下载、课程学习等过程中有任何疑问或建议,欢迎提出宝贵意见哦~我们会及时处理!
点击此处反馈
安全验证
文档复制为VIP权益,开通VIP直接复制
信息提交成功