Was:
Evaluation of different Flow Control Algorithms
Is:
Evaluation of an Enhanced Faireness
Algorithm that Avoids HOL Blocking
version 0.1
Stein Gjessing
Simula Research Lab. / U. Oslo
Oslo, NORWAY
steing@ifi.uio.no www.ifi.uio.no/~steing
3. September 2001
802-17-01-00071, sg_fcsim_01.pdf
Stein Gjessing, Simula Research Lab
An Enhanced Faireness Algorithm for
avoiding HOL blocking version
0.1
• Presented previously
• Assume Non-HOL Blocking Queues
• All stations have 2(N-1) counters that for each link
records
– the ”owner” usage of the link (sent in a control packet)
– this stations usage of the link
• Send on a link if you have used less than the ”owner”
(The ”owner” of a link is the station immediately upstream)
3. September 2001
802-17-01-00071, sg_fcsim_01.pdf
Stein Gjessing, Simula Research Lab
An RPR model in Java
• Two priority levels - Separate buffers for the two
• One tick is one byte transfer
– OC-48 is about 3ns per tick (8/3 Gbit/sec)
• 16 stations per ring
– Also run some of the experiments with 32, 64 and 128
stations – nothing special to report
• 2500 ticks cables (also run with 50 000 tick cables)
– have not noticed any difference on results from longer
cables (but have not looked close, parameters should be tuned)
• The model is a program of more than 10 lines, hence
it contains errors
~ 2 thousand lines
3. September 2001
802-17-01-00071, sg_fcsim_01.pdf
Stein Gjessing, Simula Research Lab
All experiments
• 500 byte data packet low priority, fairness
• 32 byte control packets, high priority, no fairness
• New fairness algorithm compared with
”Traditional fairness”
– ”Traditional fairness” algorithm is my implementation
of Ciscos RFC 2892
• All (except ”Random startup” at end) are stable
situations
– 99% confidence interval for mean almost not visible
• All maximum load
3. September 2001
802-17-01-00071, sg_fcsim_01.pdf
Stein Gjessing, Simula Research Lab
Hot receiver and local traffic
0
1
2
3
4
5
6
7
Hot receiver
3. September 2001
802-17-01-00071, sg_fcsim_01.pdf
Stein Gjessing, Simula Research Lab
Hot receiver and local traffic
250
250
To immediate downstream neighbor
To hot receiver (station 7)
Total received by hot receiver
200
150
200
100
50
50
0
0
0
1
2
3
4
Station number
5
6
Enhanced fairness
algorithm,
non-HOL
blocking
queues
150
Traditional fairness
algorithm,
HOL blocking queues
100
To immediate downstream neighbor
To hot receiver (station 7)
Total received by hot receiver
7
0
1
Number of packets per 100 000 ticks
3. September 2001
802-17-01-00071, sg_fcsim_01.pdf
2
3
4
5
6
7
Station number
(really per 100 160 ticks)
Stein Gjessing, Simula Research Lab
Hot receiver – no local trafic
250
250
Sent to hot receiver
Sent to hot receiver
Total received by hot
i
200
150
150
Traditional fairness
algorithm
100
50
0
0
1
2
3
4
5
Enhanced fairness
algorithm
100
50
0
Totl received by hot receiver
200
6
7
0
1
2
3
4
5
6
7
No HOL blocking - no difference
3. September 2001
802-17-01-00071, sg_fcsim_01.pdf
Stein Gjessing, Simula Research Lab
Hot sender and local traffic
0
1
Hot sender
3. September 2001
2
3
4
5
6
7
HOL Blocking an issue
in station 0 only
802-17-01-00071, sg_fcsim_01.pdf
Stein Gjessing, Simula Research Lab
Hot sender and local traffic
250
250
From immediate upstream station
From hot sender
Hot sender total output
200
150
From immediate upstream station
From hot sender
Total sent by hot sender
200
Traditional fairness
algorithm,
HOL blocking
queu
Enhanced fairness algorithm,
non-HOL blocking
queue
150
100
100
50
50
0
0
0
1
2
3
4
5
6
7
0
1
Station number
3
4
5
6
7
Station number
Number of packets per 100 000 ticks
3. September 2001
2
802-17-01-00071, sg_fcsim_01.pdf
(really pr 100 160 ticks)
Stein Gjessing, Simula Research Lab
Random traffic
• All send to all
• Fully loaded ring in both directions
– Contol and data packets compete on the wires
• Send shortest path
– when a tie , use outer ring
– Hence, outer ring is more congested, and gets
fewer packets thru
3. September 2001
802-17-01-00071, sg_fcsim_01.pdf
Stein Gjessing, Simula Research Lab
Traditional fairness
algorithm,
HOL blocking queus
80
70
60
Random traffic
Inner ring
Outer ring
70.91
70.81
55.12
55.03
80
70
50
50
40
40
30
30
20
20
10
10
0
0
1
2
99 % confidence
interval for the mean
is ~0.15 packets
3. September 2001
3
4
5
Hops away
Outer ring
70.58 70.27
60 56.23
55.21
6
7
8
Enhanced fairness
algorithm,
non-HOL blockingqueue
1
70.25
54.94
54.66
2
3
Inner ring
70.35
70.17 70.16
54.77
54.71
54.57
4
5
6
Hops away
Packets per station per million ticks
802-17-01-00071, sg_fcsim_01.pdf
70.41
54.46 54.59
7
8
99 % confidence
interval for the mean
is ~0.3 packets
Stein Gjessing, Simula Research Lab
Startup of Random traffic
- short links (remember: full load)
2500
2500 ticks long links
Enhance algorithm
Traditional algorithm
Packets per 96 000 ticks
2000
OC-48: 1500 meters
1500
1000
500
OC-48: 300 ms.
0
0
50
100
150
200
250
300
350
400
450
500
550
600
650
700
750
800
850
900
950
1000
Number of 96 000 ticks
Total number of packets received on 16 node ring
3. September 2001
802-17-01-00071, sg_fcsim_01.pdf
Stein Gjessing, Simula Research Lab
Random traffic – startup
(short links - enlarged first 1/2 of previous)
2500
2500 ticks long links
Enhanced algorithm
Traditional algorithm
Packets per 96 000 ticks
2000
OC-48: 1500 meters
1500
1000
500
OC-48: 150 ms.
OC-48: ~3 ms.
0
0
50
100
150
200
250
300
350
400
450
500
Number of 96 000 ticks
Total number of packets received on 16 node ring
3. September 2001
802-17-01-00071, sg_fcsim_01.pdf
Stein Gjessing, Simula Research Lab
Startup of Random traffic
- long links (remember: full load)
2500
2000
Packets per 96 000 ticks
50000 ticks long links
Enhanced algorithm
Traditional algorithm
OC-48: 30 000 meters
1500
1000
500
OC-48: 300 ms.
OC-48: ~5 ms.
0
0
50
100
150
200
250
300
350
400
450
500
550
600
650
700
750
800
850
900
950 1000
Number of 96 000 ticks
Total number of packets received on 16 node ring
3. September 2001
802-17-01-00071, sg_fcsim_01.pdf
Stein Gjessing, Simula Research Lab
Recorded Passthru buffer depth
- short / long links
• All Passthru buffer depth recorded at every measure
point ( each 96 000 ticks, ~1040 measure points)
• Mean:
– Traditional: 2764 / 2889 (~half of the buffers empty at any time)
– Enhanced: 21052 / 21101 (most buffers non-empty at any time)
• Max (threshold is 25 000 bytes) :
– Traditional: 25 000 / 25 500 bytes
– Enhanced: 26 000 / 25 500 bytes
3. September 2001
802-17-01-00071, sg_fcsim_01.pdf
Stein Gjessing, Simula Research Lab
Conclusion
The Enhanced Fairness Algorithm
version 0.1
• Aviods HOL blocking almost perfectly
• Is dynamic
– No pre-set loads
• Is aggressive
– Might be too agressive ?
• Simple algorithm
– Few counters
– Small control packets broadcasted
– Or can aggregate status to larger point-to-point packets
(not shown)
3. September 2001
802-17-01-00071, sg_fcsim_01.pdf
Stein Gjessing, Simula Research Lab