Transcript iSCSI Performance Experiments

iSCSI Performance
Experiments
Li Yin
EECS Department
U.C.Berkeley
Outline




Introduction
iSCSI
iSCSI Performance Challenges
Experimental Plan
Introduction

Networked Shared Storage





In contrast to direct attached storage
High availability
Better resource sharing
Easy to manage
NAS vs. SAN
Network Attached Storage
Storage Area Networks
IP SAN

Transport SCSI over IP Networks




Less cost
Longer operating distance
Leverage existing technology and management tools
Protocols to transport SCSI over IP Networks
 FCIP , iFCP ,iSCSI
iSCSI

Encapsulates SCSI command in TCP packets



SCSI commands are encapsulated into iSCSI protocol data
unit (PDU).
PDUs are sent over one or more TCP connections.
TCP as the transport protocol



Provides the reliability and congestion control method
The most dominant transport protocol
Well-supported
iSCSI Performance Challenges





TCP processing overhead
Security
Packet drops
Delay caused by long distance
…
Experiments Setup

Testbed:

Initiator:


Router:



Intel initiator
Current: NistNet controls delay, bandwidth and drop rate
Future: Router VM
Target:

DiskSim as the iSCSI target
Experimental Plan

Performance Metrics:



Throughput
I/O Rate
Three steps:



iSCSI on a separated LAN
iSCSI on a LAN with cross traffic
iSCSI on a WAN
Preliminary result


To test the impact of packet drop, delay and
request size on the overall throughput.
Setup:





Ignore the disk mechanical impact
Delay = 10ms
Bandwidth = 100Mb/s
NistNet to generate random drop
Continuously write/read (100 times) equal size
request to/from the target
Delay vs Throughput
(ms)
Drop Rate vs. Throughput
Experimental Plan (Cont.)

Issues to focus on:

How to maintain a high performance with packet
drops?

Performance drop due to



Bandwidth provision?



Bandwidth shrink
Head of Queue blocking
Might be expensive.
Router queue management mechanism with the assistance
of PNE?
Multiple TCP Connections?


Higher overall bandwidth.
Smart scheduling?
Experimental Plan (cont.)

How to achieve high performance with a long
delay?




Proxy?
Cache /Pre-fetching?
Initiator buffer?
Avoid retransmission?