Transcript iSCSI Performance - OASIS

iSCSI Performance in
Integrated LAN/SAN
Environment
Li Yin
U.C. Berkeley
Outline





Motivations
Network Congestion and Disk
Congestion differentiation
iSCSI ns-2 Simulator
Conclusion
Future work
Motivations

Storage Area Network
•
•

Hosts/Servers are connected to storage devices through
network
Advantages:
•
•
•
High availability
Better resource sharing
Easy to manage
Transport SCSI over IP Networks
• Less cost
• Longer operating distance
• Leverage existing technology and management tools
Motivations (cont.)


iSCSI is one of the most promising protocols to carry
SCSI traffic over IP networks.
SAN/LAN integration is one step towards SAN/WAN
integration
•
•
•
Utilize existing Ethernet to carry iSCSI traffic
Problem:
• iSCSI performance degrades dramatically with cross
traffic
Challenge:
• How to sustain high I/O throughput even with competing
traffic?
Why Performance Degrades?


Resources are saturated
•
First step: where is the
congestion point?
•

Cross traffic will compete for
network bandwidth, disk, CPU
and etc.
Our hypothesis: Congestion
caused by network and disk
behave differently
Second Step: how to avoid it?
Test bed Setup

Topology:
Software Router
Initiators




Target
Cisco iSCSI initiator
Target machine runs IBM iSCSI target software driver
A SCSI disk is attached to the target machine as the target
disk
Nistnet running on the software router to control delay,
packet drop rate and etc.
Experimental Setup

iSCSI target disk is attached as raw device at the
initiator
•


To filter out file system effect
Rt
Ri
16384 64KB random requests (1G data in total)
•
To reduce the cache effect
Read
Initiator: response time seen by the initiator
Ri  Tif  Tir
Tif: Initiator request finish time; Tir: Initiator request receive time

Target: response time seen by the target iscsi
driver
Rt
Ri
Rt  Ttf  Ttr
Ttf: Target request finish time; Ttr: Target request receive time
Write
Case 1: Network congestion

Use packet drops to reflect network
congestion

Nistnet introduces uniform random
packet drops

The target disk performs no other disk
operations
Experimental Results
initiator
No drop
1% drop
5% drop
target
read
write
Case 2: Disk Competition

Target machine continuously issue
random I/O requests to the target disk

No network interference is introduced
Experimental Results
initiator
No drop, No compete
4KB Competing Request
4MB Competing Requests
target
read
write
Result Analysis
initiator
No drop, No Compete
4KB Competing Request
1% Drop
target
read
write
Congestion Caused by Network


Target: How to improve performance
when the degradation is caused by
network?
Our methodology: ns-2 based iSCSI
simulator
• Simulation can help us alleviate the limitation
of available testbed configuration
• Delay
• Bandwidth
Ns-2 iSCSI simulator

iSCSI is implemented as a TCP
application.
•

all iSCSI messages are encapsulated as an
application message transmitted using TCP
Start from a simple disk models
• Parameters:
• cache hit ratio
• mean, variance, minimum processing time and
maximum processing time for various request size
Simulator Verification

16384 64KB random requests (1GB in total)
Throughput
Simulation
Testbed
Testbed
Delay
READ
WRITE
Impact of TCP Parameters

Impact of minimum timeout value (minrto)
•
Real RTT is roughly 1.5ms
Throughput
1s minrto

# of Fast Retransmission
100ms minrto
10ms minrto
# of Timeout
1ms minrto
A more precise configuration of timeout value in LAN is
useful, and is possible
Conclusions




iSCSI performance in LAN/SAN integrated
environment is very important
Performance degradation caused by network
and disk is possible to be distinguished
Ns-2 based simulator can match the real curve
closely
TCP parameter configuration have big impact
on the performance
Future Works

Congestion point detection

Simulation:
• better monitoring point?
• more complicated cross traffic
•
•
•
•
More verification on write operations
Bursty Drop Model
More realistic system and disk model
More analysis on how TCP will affect the
performance?
Future Work (Cont.)

Design scheme to improve performance when congestion
happens in network
• Can we avoid packet drop?
•
•
•
•
•
Resource management?
Shift to another target?
•
•
Read: Multiple replica may exist
Write: write to some target temporarily, re-organize data later
QoS management
Avoid iSCSI drop rate below some threshold
If we have to, which packet to drop?
•
Leverage application layer information
•
•
•
Application importance
iSCSI header packet vs. iSCSI data packet?
Packets belong to same iSCSI data unit or multiple data units?