Introduction
100Mbit
1Gbit
Conclusion
1/10
Linux’s DCCP—slower than RFC 1149
Andrea Bittau
June 21, 2006
DCCP
Introduction
100Mbit
1Gbit
Conclusion
2/10
DCCP is a protocol with modular congestion control algorithms:
CCID2: TCP-like. Use all available bandwidth but high
variance in transmission rate.
CCID3: TFRC. Low variance in transmission rate by
sacrificing some bandwidth.
Relevance:
Define a CCID which suits application needs.
Unreliable: do not waste time retransmitting—send new data.
Linux implementation status
Introduction
100Mbit
1Gbit
Conclusion
3/10
Status of DCCP implementation in Linux 2.6.17 (latest):
Has CCID2 and feature negotiation: thanks to this project.
Most of the DCCP spec is implemented.
Unstable and buggy. Wait for next Kernel before sending
flames.
Now that we got a lot of code in Linux, lets make it work.
Performance tests
Introduction
100Mbit
1Gbit
Conclusion
4/10
The next milestone is to get DCCP to be stable and fast.
Stress efficiency.
100Mbit/s, 0ms RTT.
1Gbit/s, 0ms RTT.
Hardware:
Athlon 1.2GHz.
PCI 33MHz 32-bit. (≈1Gb/s.)
Realtek 8169 1Gb/s.
Stress correctness.
100Mbit/s, 100ms RTT.
1Gbit/s, 100ms RTT.
100Mbit/s, 0ms RTT
Introduction
100Mbit
1Gbit
Conclusion
5/10
Throughput: 94Mbit/s. Success!
500
cwnd
pipe
packets
400
300
200
100
0
0
10
20
30
time (s)
CCID2 reported 45 losses:
37 were incorrect IP lengths.
8 were incorrect DCCP checksums.
40
50
60
100Mbit/s, 100ms RTT
Introduction
100Mbit
1Gbit
Conclusion
Simulating delay with netem
6/10
Average throughput: 15.6Mbit/s. Failure!
500
cwnd
pipe
packets
400
300
200
100
0
0
10
20
30
40
50
60
70
time (s)
CCID2 reports 10 losses. They all were incorrect IP lengths.
TCP averages 75Mbit/s in this scenario.
1Gbit/s, 0ms RTT
Introduction
100Mbit
1Gbit
Conclusion
7/10
Throughput (CPU limited at TX):
UDP: 335Mbit/s.
TCP: 310Mbit/s.
DCCP: 230Mbit/s.
Kernel profile for DCCP run
samples
3734
769
362
357
230
182
...
151
%
20.1185
4.1433
1.9504
1.9235
1.2392
0.9806
0.8136
image name
vmlinux
vmlinux
vmlinux
vmlinux
vmlinux
vmlinux
symbol name
handle_IRQ_event
__do_softirq
rtl8169_rx_interrupt
rtl8169_interrupt
__copy_from_user_ll
rtl8169_poll
dccp_ccid2.ko
ccid2_hc_tx_packet_recv
1Gbit/s, 100ms RTT
Introduction
100Mbit
1Gbit
Conclusion
8/10
Average throughput: 75Mbit/s. Failure!
3000
cwnd
pipe
packets
2500
2000
1500
1000
500
0
0
10
20
30
time (s)
40
50
60
CCID2 reports 543 losses. Some packets were never received,
others were dropped in receiver’s DCCP stack.
TCP achieves 206Mbit/s (CPU limited at TX).
Summary
Introduction
100Mbit
1Gbit
Conclusion
9/10
Link speed
100Mbit/s
100Mbit/s
1Gbit/s
1Gbit/s
RTT
0ms
100ms
0ms
100ms
Throughput
94Mbit/s
15.6Mbit/s
230Mbit/s
75Mbit/s
Bottleneck
Link
Implementation?
CPU
Implementation?
I have also seen CCID2 do line rate with a powerful box on
1Gbit/s, 0ms RTT.
Conclusion
Introduction
100Mbit
1Gbit
Conclusion
10/10
Preliminary results are not that bad:
We expected more blue screens rather than numbers.
Future work:
[Get two powerful boxes.]
Achieve 1Gb/s on a large delay network.
Devise a performant API. Currently, one system call transmits
one packet. Maybe we need something like writev.
Implement a CCID which meets the requirements of VLBI.