Title: Max Min Fairness
1Max Min Fairness
- How define fairness?
- Any session is entitled to as much network use
as is any other - .unless some sessions can use more without
hurting others - Other definitions
- Network usage depends on the resource consumption
by the session - Pay/bid for what you use
2A Simple Example
- Max-min allocation 1/3, 1/3, 1/3, 2/3
3How to Calculate Max Min Flow Share
- Fluid model
- Increase the flow until some pipe fills-up.
- Fix the bandwidth of the bottleneck flows
- Continue with the unfixed flows
- Can be done efficiently by calculating the
bottleneck link at step 1
4Buffer management and admission control
- Simplest admission policy
- Accept packets until buffer is full (tail drop)
- We already saw
- Tail drop is not kind to TCP flows
- RED can be used to avoid tail drop
5Reminder Hallelujah for RED
- Random early detection (RED) makes three
improvements - Metric is moving average of queue lengths
- small bursts pass through unharmed
- only affects sustained overloads
- Packet drop probability is a function of mean
queue length - prevents severe reaction to mild overload
- Can mark packets instead of dropping them
- allows sources to detect network state without
losses - RED improves performance of a network of
cooperating TCP sources - No bias against bursty sources
- Controls queue length regardless of endpoint
cooperation
ECN
6How does it work?
1
7So problem is solved?
- Fairly easy to implement in hardware!
- Can work in wire-speed!
- All we need to do is set the parameters..right ?
- Turns out there is no universal good set of
parameters - Some studies show RED has NO advantage over tail
drop. WHY?
8parameters
- avgQ (1-wq)avgQwqq
- Floyd-Jacobson
- Wq 0.002, not less than 0.001
- max_p 1/50,
- max_th at least twice min_th
- max_th-min_th larger than the q increase in RTT
- Future work ..
9So does it help us to surf?
- Tuning RED for Web Traffic, Christiansen et al.,
SIGCOMM 2000 - compared to a (properly configured) FIFO queue,
RED has a minimal effect on HTTP response times
for offered loads up to 90 of link capacity, - response times at loads in this range are not
substantially effected by RED control parameters,
- between 90 and 100 load, RED can be carefully
tuned to yield performance somewhat superior to
FIFO, however, response times are quite sensitive
to the actual RED parameter values selected, and - in such congested networks, RED parameters that
provide the best link utilization produce poorer
response times.
10SPRINT study (Diot et al.)
- A parallel study, presented at NANOG 2000
- Testbed
- with CISCO routers (7500)
- with Dummynet
- used recommended RED and GRED parameters
- Heterogeneous delays (120 to 180 ms)
11Traffic characteristics
- 16 to 256 TCP connections sharing the bottleneck.
- Experimental traffic generated by Chariot
- long-lived TCP connections.
- more realistic traffic mix
- 90 short lived TCP connections (up to 20
packets) - 10 long lived TCP connections
- 1Mbps UDP in both cases
12Testbed (CISCO routers)
7500
7500
10 Megs
13Testbed (Dummynet)
7500
7500
10 Megs
Dummy net
100 Megs
14What is Dummynet?
application
dummynet
network
15Metrics observed
- Aggregate goodput through a router
- TCP and UDP loss rate
- Consecutive losses
- Queuing behavior
16Aggregate goodput (long-lived TCP)
17256 short and long lived TCP connections
18Consecutive packet losses (long lived)
19if we use optimal RED parameters
20Consecutive packet losses (realistic traffic mix)
21Queuing behavior (256 long lived connections)
22Queuing behavior (256 connections, realistic mix)
23Diots summary
- No significant difference on goodput, TCP losses
and UDP losses. - On consecutive losses, clear advantage to GRED
and GRED-I. - gentle modification solves many RED problems.
- Oscillations no clear winner. Traffic seems to
be the determining factor.
24From the ISP standpoint ...
- Not clear there is an advantage in deploying RED,
GRED, or GRED-I. - Maybe GRED-I is an option if one can find a
universal exponential dropping function. - ECN will work with any scheme.
- Not clear the solution is in the AQM space.
25GRED-I with exponential dropping function
1
buffer size
26About Fair Queuing ...
- Not only feasible easy at the edges!
- www.agere.com (an example)
- vendors support from 64k to 200k flows
- Really fair
- everybody gets what he/she paid for
- local signaling (end host to CPE)
27fair queueing at the edge
- Core-stateless fair queueing
- WFQ is hard to do at the core
- Edge routers estimate rate and label packets
- Core routers maintain FIFO queues and drop based
on label
28(No Transcript)
29CSFQ summary
- Better than FIFO and RED
- Similar to FRED
- Not as good as DRR
30Rainbow fair queueing
- Similar to CSFQ
- Have similar performance as CSFQ
- Enable applications to mark packets and achieve
better goodput
31Rainbow Fair Queueing (RFQ)
- Example
- A 10 Kbps B 6 Kbps C 8 Kbps
- Each layer 2 Kbps
32RFQ basic mechanism
- (1) the estimation of the flow arrival rate at
the edge routers - (2) the selection of the rates for each color
- (3) the assignment of colors to packets
- (4) the core router algorithm
33Rainbow Fair Queueing (RFQ)
- (1) the estimation of the flow arrival rate at
the edge routers
- rinew arrival rate
- tik arrival time of flow I
- lik length of the kth packet of flow I
- K a constant
- Tik tik tik-1
34Rainbow Fair Queueing (RFQ)
- (2) the selection of the rates for the rates for
each color
- ci i color average rate of packets
- N total number of colors and multiple of b
- a,b determine the block structure
- P the maximum flow rate in the network
35Rainbow Fair Queueing (RFQ) Example
c0 c1 c2 c3 c4 c5
c6 c7 P/16 P/16 P/16 P/16 P/8
P/8 P/4 P/4
36Rainbow Fair Queueing (RFQ)
- (3) the assignment of colors to packets
- Suppose the current estimate of the flow arrival
rate is r, and j is the smallest value satisfying
. - Then the current packet is assigned color
with probability .
37- (4) the core router algo.
- Conditions to decrease color
- q threshold
- Flow bw
- Positive gradient
- Hold you horses
- Conditions to increase color
- Time
- Flow below service rate
38Rainbow Fair Queueing (RFQ)
- Weighted RFQ
- wi weight for flow i
- cj wicj
39Simulations A single congested link
40Fairness flow i sends at 0.313i
41Throughput TCP flow
42Throughput UDP flows
43Control Responsiveness10Mbps 8x1M?7x1M8M
44Simulations Performance Effects of Buffer Size
45Simulations TCP Performance Under Various round
Trip Delay