iSCSI Extensions for RDMA: A Performance Enhancing Protocol Study - CSDN文库

PDF文件

下载需积分: 9 | 275KB | 更新于2025-01-21 | 149 浏览量 | 举报收藏

立即下载

"本文深入研究了iSCSI（Internet Small Computer System Interface）协议与RDMA（Remote Direct Memory Access）技术的结合，即iSCSI Extensions for RDMA (iSER)。iSCSI是一种将SCSI（Small Computer System Interface）应用协议映射到TCP/IP上的IETF标准，旨在实现存储流量在标准TCP/IP网络上的融合。高效地在TCP/IP网络上传输和放置数据对于这种融合至关重要。iWARP协议套件提供了在TCP/IP网络上实现RDMA语义的能力，允许通过IP架构进行高效的内存到内存数据传输。文章详细探讨了iSER协议的设计过程，该协议将iSCSI协议映射到iWARP协议栈上。通过这项研究，作者揭示了iSER如何利用通用RDMA硬件实现iSCSI的数据高效移动，并对iWARP架构特性进行了讨论，这些特性是在设计iSER时考虑的关键因素。这包括对低延迟、高带宽以及减少CPU干预的需求，以优化存储操作的性能。此外，iSER的优势在于它能够减少协议处理的开销，因为它直接在内存级别进行数据传输，从而减轻了主机CPU的负担。这不仅提高了数据传输速度，还改善了整体系统效率。通过利用iWARP的特性，如零拷贝（zero-copy）传输和远程直接写入（remote direct write），iSER能够在保持TCP/IP的兼容性和可扩展性的同时，提供接近于光纤通道（Fibre Channel）的速度。文章可能还涵盖了iSER的实现挑战，如兼容性问题、安全性和管理复杂性，以及如何通过iWARP的流控制和拥塞控制机制来保证网络的稳定性和可靠性。同时，可能也讨论了iSER相对于其他存储协议（如FCoE，Fibre Channel over Ethernet）的竞争优势，以及在云计算、大数据分析和高性能计算等场景中的应用潜力。这篇论文对理解iSCSI与RDMA的集成，以及如何通过iSER提升存储网络性能提供了深入的技术见解，对于网络存储和数据中心架构的设计者和管理员来说具有很高的参考价值。"

reassembly buffer requirements for TCP depending on

framing solutions adopted. The draft presents analysis

demonstrating that MPA allows reassembly buffer size to

scale as a function of maximum segment size rather than as

a function of bandwidth delay product. Another important

b

enefit of DDP/MPA is that it ensures the presence of an

iSCSI PDU (or more generally, any upper layer protocol

PDU) at a fixed offset in the first DDP segment of a DDP

Message. If the location of the iSCSI PDU had no

relationship to the TCP segment boundaries, the receiver

complexity is largely increased. The “Analysis of MPA

over TCP Operations” [12] also provides an illustration for

the benefit of header alignment in streamlining the DMA

operations and computational load of the receiver.

The iWARP protocol suite is implemented in an RNIC

which is capable of providing the direct data placement

function without requiring firmware or hardware

customizations in the RNIC for each ULP. The iWARP

protocol suite is depicted in Figure 1.

Internet Protocol (IP)

Transmission Control

Protocol (TCP)

Markers with PDU Alignment

(MPA)

Direct Data Placement

Protocol (DDP)

RDMA Protocol (RDMAP)

iWARP

Figure 1 iWARP protocol suite over TCP/IP

The primary unit of information exchange on an iWARP

connection is a DDP Segment that is self-describing from a

data placement perspective. The MPA framing support is

used by the RNIC to locate the DDP control information of

the current DDP segment in the TCP stream, whether or not

the TCP packets arrive in order. The DDP control

information supports the direct data placement and enables

the RNIC to steer the payload to its final memory location.

The RDMA support is provided in the RDMAP control

information and it enables the ULP to specify RDMA Read,

RDMA Write, or Send semantics to transfer the data.

A ULP interacts with an RNIC via the RNIC Interface (RI)

using a set of functional interfaces called “Verbs” [9].

From this perspective, the RI is also called a “Verbs

provider” and the ULP a “Verbs consumer”. Verbs,

covering various functional aspects of the RI, are defined in

the iWARP architecture. The Verbs consumer accesses the

RI by creating one or more Queue Pairs (QPs), each of

w

hich consists of two Work Queues (WQs): a Send Queue

(SQ) and a Receive Queue (RQ) as shown in Figure 2. This

QP is associated with a TCP connection (one to one

mapping) by the Verbs consumer for carrying out the send

and receive operations. Each request to the RI by the

consumer takes the form of a Work Request (WR) that the

consumer posts to the SQ or RQ as appropriate by invoking

Verbs in order to convey the request to the RI. All

outbound RDMA operations such as RDMA Read, RDMA

Write, and Send are initiated via WRs posted to the SQ.

All inbound solicited data carried by the RDMA Write and

RDMA Read Response messages is stored by the RNIC

directly into the ULP buffer(s). The control and un-solicited

data operations such as receiving incoming Send Messages

are satisfied via Work Requests posted to the RQ. Each

WQ (SQ or RQ) is associated with a Completion Queue

(CQ) that notifies the Consumer of the completion of the

requested operation via a Work Completion (WC).

Verbs

Consumer

Verbs

Provider

RI

SQ RQ

CQ

CQ

WR

WR

QP

WC WC

Legend

WR – Work Request

WC- Work Completion

RI – RNIC Interface

SQ – Send Queue

RQ – Receive Queue

CQ – Completion Queue

QP – Queue Pair

Figure 2 iWARP functional interface

As RNICs support a diversified set of applications, the use

of RNICs promotes the convergence on the server hardware

and networking requirements. As such it creates larger

economic thrust for creating a larger RNIC market than

several application-specific markets for iSCSI NICs, IPC

NICs, TOE NICs etc. An RNIC employs the iWARP

protocol suite defined by the RDMA Consortium that have

been accepted as drafts by the IETF Remote Direct Data

Placement Working Group (RDDP WG) and are expected

to be supported by a broad list of hardware and software

vendors.

Proceedings ot the ACM SIGCOMM 2003 Workshops 211 August 2003

剩余10页未读，继续阅读

chenchihweijack

粉丝: 5

最新资源