Title: Michael MelliarSmith
1AFOSR Multidisciplinary University Research
ProgramProtocol Engineering Research Center
Performance ofMulticast Protocols for CORBA
Michael
Melliar-SmithDepartment of Electrical
Computer EngineeringUniversity of California,
Santa BarbaraJon Sjogren, Program Manager
2Group Multicast Protocols
- Typical group multicast protocols are designed
for - Complex applications
- Many concurrent activities
- High message traffic
- Performance is determined by time spent in queues
- Time spent in queues is determined by throughput
- Our Totem protocol is designed for throughput
- Flow control
- No collisions on the Ethernet, allows high
utilization - Collects several messages to transmit in a batch
- Must wait for the token keep the ring small
3Embedded Applications
- Typical embedded applications
- Simple with fewer messages
- Impose low load on the network
- Message delivery latency is critical
- Message order is determined by priorities
- Totem is not designed for embedded applications
- Waiting for the token increases the latency
- Extra Totem process increases the latency
- Message order ignores priorities
4Experimental Testbed
- Pluggable FT Infrastructure with Totem
- Three-way replicated server, unreplicated client
- 1KByte payload in request and reply messages
- Four Pentium III PCs, single 1GHz CPU, 256 RAM,
Linux Mandrake 7.2 OS - 100 Mbps FastEthernet LAN
- eOrb from Vertel
5Latency Measurements
Latency in ?seconds
6Latency for a Simple Request/Reply
7Sending Delay
Semi-Active Replication
8Fault Detection and Recovery Time
- Crash fault detection time 3 ms
- Recovery time 100 ms
9Performance
Three-Tier Application
Stores Records
for Account A
Account A
G1
G2
Gateway
Gateway
AccountManager
OTS Server
Backup
Backup
Backup
Backup
Primary
Primary
Stores Records
for Account B
Account B
Client
Application Server
Database
10End-to-End Performance Three-Tier Application
Latency
Throughput
11Performance
Three-Tier Application
Application Logic Processing Time
2PC Processing Time
12Summary
- Group multicast protocols, such as our
Totem protocol, are designed for complex
applications and achieve high throughput - For applications that require low latency,
such as real-time embedded applications,
a different
approach is required