Improving Loss Resilience with Multi-Radio Diversity in Wireless Networks

1

• Improving Loss Resilience with Multi-Radio
  Diversity in Wireless Networks
• by Allen Miu, Hari Balakrishnan and C.E. Koksal
• Appeared in ACM MOBICOM 2005, was considered as a
  candidate for the best paper award.

2
What is the paper about?

• The paper looks at multi-radio diversity and uses
  the fact that the fading experienced at the
  different radios are different to improve
  performance in WLANs.
• The major contribution of this paper is that it
  proposes a new technique called Frame Combining
  using which, it tries to combine two frames
  received by the two radios, in error, to
  reconstruct the transmitted frame.
• The effort is thorough the authors do an
  implementation of their proposed approach,
  provide some analytical insights and also do
  simulations to study the problem deeper.

3
What did I learn from this paper?

• Obtained insights into how much, and why,
  multi-radio diversity can help improve
  performance.
• How does any kind of diversity affect rate
  control ? (to some degree)
• Some interesting interactions between the
  physical, MAC and network layers which I will
  highlight in the presentation.

4
The MRD System

• The idea is to use multiple radios or multiple
  APs in a wireless LAN to simultaneously receive
  transmitted frames.
• Why does this help ? -- The radios are not
  spatially collocated.
• Thus, the wireless channel to the two radios
  differ (path diversity)
• Thus the errors experienced at one radio, would
  differ from that at the other.
• So, you would get two copies of a transmitted
  frame, but possibly at errors at different
  locations.
• Of course, if one of the frames is error free,
  that can be delivered to the IP layer (selection
  diversity).
• If not, can we combine the corrupted frames ?

5
A High level view

• Note Unless I state otherwise, all figures that
  I have are from the paper itself.

• Of the two APs, one is called an active AP -- it
  is in communication with the client.
• The others passively listen and try to gather
  frames.

6
What are the challenges ?

• How does this frame combining (done at a layer
  that sits above the MAC) interact with the 802.11
  MAC and the PHY layer ?
• Specifically, the frame combining make take some
  time, and so how can you acknowledge and provide
  retransmissions of reconstructed or salvaged
  frames ?
• How does this frame recombining work with the
  auto rate control ?
• How are the bit errors distributed ? Can frames
  even be salvaged and if so how ?
• The paper tries to address these questions while
  proposing a new technique for frame combining.

7
Frame Combining

• The idea is simple. Divide each frame into NB
  blocks, each of fixed size (the last block may
  possibly have fewer bits).
• Let us say we want to see if we can reconstruct
  the frame from two copies that are received.
• Clearly if any of the copies is ok (CRC is ok),
  then, the packet is successful.
• We look at those blocks that differ -- as an
  example, the ith block of the first copy might
  differ from that of the second copy.
• Assemble a combined frame with different
  possibilities --choosing different blocks from
  either of the copies. If CRC passes, then a
  success.

8
More about frame combining

• What should the block size be ?
• Note that the previous combining method is
  simple, but its running time is exponential in
  terms of the number of differing blocks D.
• If you have two copies, then you need about 2D
  CRC check operations.
• So, clearly you want to keep D small which means
  that you may want to reduce the total number of
  blocks i.e., increase B, the number of bits per
  block.
• However, if you do this, then, the possibility of
  successfully recombining reduces (Why ?).

9
Analyzing Frame Combining

• Let the frame combining failure probability be
  pf.
• Let there be a bit error model characterized by
  bursts of b bit errors.
• pf is the fraction of frames that cannot be
  corrected with combining out of those that could
  not be corrected by the soft selection.
• The assumption made is that the loss rates
  observed at the two receivers are independent of
  each other. -- The paper corroborates this claim
  by experimental results.
• The errors are clustered and occur with a
  periodicity.

10
Bit Errors

• The authors claim that the error pattern is in
  line with what is used.
• QAM-64 modulation on OFDM with a rate 2/3 code.
• With QAM-64, you transmit 6 bits/symbol.
• This means, that for each transmission on 50 OFDM
  carriers, you have 50 x 6 300 bits.
• Since you use a rate 2/3 code, you decode 3
  symbols at a given time -- each carrier carries 3
  symbols.
• Thus, you have approximately 900 bit transmission
  patterns on the different carriers that repeat.
• Since each carrier is likely to experience
  similar fades periodically (static), the error
  distribution repeats about 1000 s.

11
Assumptions made for computing pf

• In order to compute the frame combining failure
  probability pf, the authors make the following
  assumptions
• The burst of errors is of fixed to b bits.
• The number of bits per block B is much larger
  than b and thus
• the probability that two blocks have errors that
  overlap is negligible.
• they ignore the possibility that the errors can
  spread over more than one block -- i.e., the
  errors are completely contained within a block.
• the number of burst errors that a block can hold
  are not fixed.
• Note that given that b 300 bits and a block is
  more than say 200 bytes, these are reasonable
  assumptions.

12
Notations and some details

• Db,i -- number of b-bit sequences with errors in
  a given frame received at receiver Ri.
• N1 and N2 represent the sets of blocks that
  contain errors in frames received at receivers 1
  and 2.
• Note that two receivers are considered.
• Then, the intersection of the two sets N1 and N2
  represent those blocks that have errors in both
  frames.
• Now, if this intersection (N1 ?N2 ), contains no
  errors, then, it means that the frame can be
  decoded.

13
Computing pf

• First, assume that bit error sequences (b bits in
  error) occur uniformly over the frame.
• Let frame 1 have d1 errors and frame 2 have d2
  errors.
• Then,

Why is this true ?
14

• This is similar to the problem wherein we have
  NB buckets, and d1 balls we put the balls
  (probably more than NB) into those NB buckets.
• We want to compute the number of ways in which
  we can put balls into these buckets.
• Note that some buckets may have multiple balls
  and so we can have empty buckets.

15

• So, we have the first NB buckets, and in addition
  have (d1-1) dummy buckets. Note that there is at
  least one bucket which contains balls (all
  balls).
• If a ball falls into a non-empty bucket, we put
  the ball into one of the dummy buckets.
• Thus, the total ways we can do what we want is to
  choose d1 buckets out of the NB d1 -1 buckets.

16
Given this....

• We remove the conditioning to get

where

• Note This is an upper bound on the frame
  combining failure probability.

17
Looking at pf

• Note that if the burst error size is small,
  errors are more uniform, and even for large NB,
  probability of combining successfully is small.
• With bursty errors (as observed), pf gets lower
  with NB.
• But beyond a certain point, increasing NB does
  not help much.

18
What does this mean?

• It is necessary to keep NB small, so as to reduce
  complexity.
• So, NB can be set to a small value (6-10) and
  still performance is ok.

19
Retransmissions

• Clearly link layer ACKs can result in erroneous
  conclusions.
• So the authors disable link layer
  retransmissions.
• Retransmissions are always invoked by the MRD
  layer.
• If packet reception is successful, synchronous
  ACK is received (at link layer).
• Else, this indicates either a frame with errors
  or an ACK failure.
• A frame with errors is either recovered using
  soft decision or frame combining if this fails,
  the frame may be stored with the hope of trying
  to combine with later retransmitted versions.
• The sender expects an ACK sometime in the future.
  If prior to a time-out no ACK is forthcoming, it
  can request for an ACK -- to explicitly denote
  success or a failure.
• This is called the RFA (request for ACK).

20
RFA

• The RFA needs to explicitly state which frame is
  in question.
• Use of a flag in the frame header to indicate an
  RFA.
• When an link layer ACK fails, the MRDS (sender)
  simply stores the packet and proceeds with
  subsequent transmissions.
• It can perform upto a certain number (N) of
  future transmissions (from the first unACKed
  frame).
• Frame removed from buffer after K
  retransmissions.
• If the MRD-ACK indicates a frame recovery
  failure, the frame is retransmitted.
• If no MRD-ACK (higher layer) is received,
  retransmission after a time-out.

21
Link layer ACKs ?

• Why did they not disable Link Layer ACKs ?
• Needed for carrier sensing (virtual).
• Second, the synchronous ACKs have already a
  reserved channel.
• Loss is less probable.
• MRDS ACKs, on the other hand, need to contend for
  the channel. So they may be either lost or
  delayed.

22
Rate Adaptation

• With autorate or rate adaptation, the data rate
  is lowered if loss is encountered and increased
  with successful packet delivery.
• Not good with MRD -- not all radio receptions
  taken into account.
• The authors implement their protocol on the
  Atheros 5212 chipset driven by the Multiband
  Atheros driver.
• The authors modify the driver to make it fit with
  the MRD implementation.

23
How ?

• The original driver MADWIFI -- invokes TXCALLBACK
  to update numtx and numtxok after each frame
  transmission.
• Rate is adjusted every T seconds.
• If frame delivery rate is above 90 for S
  consecutive observation periods, then increase
  bit rate.
• If frame delivery rate drops below threshold D,
  then reduce rate.
• The authors introduce a new function
  MRD_CALLBACK.
• Very simple fix -- count the number of
  transmissions ACKED by MRD_ACKs.
• Note that MRD_ACKs are cumulative -- so they have
  a clear picture of what was received even if some
  were lost.

24
The code
The change
Original Code of MADWIFI
25
Other details

• The value of D is also fixed in the original
  spec.
• The authors argue that with reduced rate, the
  throughput obtained should be the same as with
  the higher rate.
• Lower rate but higher reliability.
• Thus, they claim that the fixed value of D (0.5)
  is too low (based on this argument) and change
  the value of this. (Leave it to you to read more).

26
Implementation

• Pentium PCs, Linux Kernel 2.4.20, 802.11 a/b/g
  wireless interfaces based on the Atheros 5212
  chipset.
• They have modified the MADWiFi driver.
• Simple experiments -- they set retry limit to
  zero at the MAC layer (so no retransmissions
  there).
• This however, disables CSMA backoff and they say
  that they will look at it in the future.
• Some discussion on the implications.
• Since MRD-ACKs are not ACKed, they are
  transmitted in broadcast mode.

27
Implementation (contd).

• CTX -- Combiner Transmit header.
• NTX -- number of attempted transmissions, 1 RFA
  bit to indicate that the sender has pending
  frames.
• seq -seq number.
• useq -- oldest transmitted data frame in MRDS
  that has not been ACKed.

28
Identifying combinable frames

• Look at the MAC layer source address and seq
  number in CTX to identify copies of the same
  network layer frame.
• Since MRDS has to correctly identify the frames,
  a 4 byte CRC protection is used to protect the
  CRC and CTX headers.
• If either of these headers are corrupted, the
  frame is dropped.

29
MRDS-ACK Implementation

• Magic value distinguishes the MRD-ACK packet from
  other downlink data payload.
• N bit transmit state -- indicates the success
  failure of up to N consecutive frames.
• seq number -- seq value of the first data frame
  in the bit vector being acknowledged.
• Link layer checksum used to detect errors in this
  packet.

30
Experiments

• The authors perform experiments with low
  variability (LOVAR) -- where client is static and
  high variability (HIVAR) where the client is
  mobile.
• Two APs -- one is the Master and the other is the
  Monitor.
• The client is run in the 802.11 Managed mode
  (i.e., it is in the LAN access config.).
• They pick NB (number of blocks) to be six. (look
  at discussion on complexity tradeoffs).

31
Setup for HIVAR experiments
32
Some Results

• Note that only a small fraction of frames
  recovered using the combining process.
• Throughput with just R1 or R2 were 8.25 and 6.42
  Mbps.
• With MRD-R1 and MRD-R2, the average throughputs
  were 18.7 and 18.36 Mbps.
• Still lower than 31 Mbps UDP throughput (computed
  theoretically).

33
Failure of frame combining

• They do thorough simulations to see why this
  happens.
• They argue that NB was too small.
• They also look at complexity versus efficiency
  trade-offs that I will not discuss here.

34
Rest..

• I wont discuss the rest of the paper but
  hopefully, this has shown what it contains.
• The final set is on LOVAR experiments, what
  happens there and finally, some discussion on
  back-offs -- if CSMA backoffs were invoked, then
  they wrongly cause nodes to back off
• The packets may still be recovered.