EQUIP Connection JCP

1
EQUIP Connection JCP

• Chris Greenhalgh (Jim Purbrick)
• University of Nottingham

2
Contents

• TCP
• Qualities
• Problems and issues
• UDP characteristics
• JCP
• Design
• Responses to problems and issues
• Availability and experience
• Plans

3
Qualities of TCP

• Simple semantics
• easy to use
• Reliable
• Totally ordered
• Robust
• Flow control
• wont swamp the receiver
• Congestion control
• wont swamp the network
• shares bandwidth reasonably fairly
• Widely available

4
Problems with TCP 1 Delays from NAGLEs
algorithm
Can be Disabled
app
send
recv
app
But see also flow and congestion control
send-buf
1
2
5
Problems with TCP 2 Delays from packet loss
app
send-buf
send
recv
app
recv-buf
1
6
Problems with TCP 3 Limited buffer management
app
send-buf
send
recv
app
1
7
Other issues

• Byte stream
• No message boundaries
• Packets do no correspond to application units
• Large send buffer
• Needed for performance on high RTT networks
• But makes problem 3 worse
• No priority system (since totally ordered)
• No connection probing
• Requires more ambiguous application-level
  keep-alives
• Limited control over protocol timeouts (e.g.
  host)
• No connection quality reporting

8
UDP

• Message-based ?
• Limited message size ?
• No flow or congestion control ?
• Unreliable ?
• Different API and style of use ?

9
Jims Communication Protocol Design overview

• Unicast, implemented over UDP
• TCP-based flow and congestion control
• Robust, fair
• Message-oriented (rather than byte stream)
• Per-message choice of (un)reliable delivery
• Multiple priority levels
• Totally ordered within each priority level
• Implements the same connection-oriented equip.net
  abstraction API as TCP Connection
• Easy to use runtime choice of protocol

10
JCP Delays due to NAGLEs algorithm

• Resolution
• No NAGLEs algorithm ?
• Packet-based (rather than byte-based) flow and
  congestion control windows, so packet and byte
  count both bounded
• C.f. TCP without NAGLE, (up to one packet per
  byte in the window)

11
TCP Recap Delays from packet loss
app
send-buf
send
recv
app
recv-buf
Lost!
1
1
1
1
2
2
1
2
2
RTT
2
2
ACK.1
2
1
1
1
1
2
1
1
2
1
2
ACK.3
2
12
JCP Delays from packet loss

• Resolution
• 1 unreliable
• Sender discards 1 on sending
• Receiver delivers 2 immediately and ACKs
• 1 lower priority than 2
• Receiver delivers 2 immediately
• Resending and delivery of 1 as for TCP
• Limitation
• 1 reliable and same priority as 2
• Same as TCP

13
TCP Recap 3 Limited buffer management
app
send-buf
send
recv
app
1
1
1
1
2
2
1
NAGLE, flow or congestion window
1
1
3
3
2
1
1 and 2superceded
ACK.2
3
2
1
2
3
2
3
2
3
2
3
1 and 2still delivered
2
3
3
ACK.4
2
3
14
JCP Limited buffer management

• Resolution
• Messages which are queued but not actually sent
  can be cancelled (e.g. 2 above)
• Messages can be sent with higher priorities
  causing early-as-possible insertion into the
  sending window
• Limitation
• Messages which have actually been sent reliably
  must still be delivered reliably
• Messages which have actually been send unreliably
  must still be ACKed for flow and congestion
  control

15
JCP Other issues

• Message-based
• Preserves message boundaries
• Network packets correspond to application units
  (important for unreliability and priority)
• Large send buffer
• Still needed for performance on high RTT networks
  with reliable data
• Only data actually in the sending window is
  committed
• Simple priority system implemented
• Includes connection probing (0.2 Hz default)
• Per-connection control over protocol timeouts
• Some connection quality reporting (limited?)

16
JCP Limitations

• User-space implementation (less responsive)
• TCP-like ACK-based flow and congestion control
• Gives bursty behaviour in the face of losses
  (compare rate-based control)
• Packet loss is assumed to signal congestion
• Like TCP, and most IP work
• Not always true with wireless networks
• No forward error correction (FEC) capability
• Priority scheme unvalidated
• Very simple expression of (un)reliability
• Future options max. retry time, max. resend
  count,

17
JCP Availability and Experience

• Availability
• EQUIP module/package equip.net
• C and Java implementations
• EQUIP dataspaces
• equipu URL scheme
• Experience
• Bystander August2002
• Successful partially connected tests
• Ian anecdotal improvement

18
JCP Plans

• Experimental comparison
• EQUIP dataspace over TCP
• EQUIP dataspace over JCP
• IBM TSpaces
• Local WaveLAN tests analysis
• WaveLAN quality vs loss rates
• JCP characteristics and responsiveness
• Validation and analysis