CA2361687A1 - Red ghosts virtually eliminate delayed reactionto packet loss - Google Patents
Red ghosts virtually eliminate delayed reactionto packet loss Download PDFInfo
- Publication number
- CA2361687A1 CA2361687A1 CA002361687A CA2361687A CA2361687A1 CA 2361687 A1 CA2361687 A1 CA 2361687A1 CA 002361687 A CA002361687 A CA 002361687A CA 2361687 A CA2361687 A CA 2361687A CA 2361687 A1 CA2361687 A1 CA 2361687A1
- Authority
- CA
- Canada
- Prior art keywords
- ghost
- packet
- list
- discarded
- visibility
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Abandoned
Links
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L47/00—Traffic control in data switching networks
- H04L47/10—Flow control; Congestion control
- H04L47/32—Flow control; Congestion control by discarding or delaying data units, e.g. packets or frames
Landscapes
- Engineering & Computer Science (AREA)
- Computer Networks & Wireless Communication (AREA)
- Signal Processing (AREA)
- Data Exchanges In Wide-Area Networks (AREA)
Abstract
The method for managing multiple flows through a communications de-vice includes maintaining a list of "ghost" flows. Any packet subject to dis-card due to an active queue management scheme like RED (Random Early Detection) is first checked against the ghost list. A non-ghost packet is discarded and a new entry is created in the ghost list. Ghost packets are discarded in proportion to their visibility, initially 1/2. The flow is removed from the list after one round trip time (perhaps when the lost packet is re-transmitted). The ghost list compensates for the delay between the time a packet is discarded and the time the packet's source receives notification of this loss through duplicate acknowledgements and consequently halves its transmission window. This reduces excess packet loss and resulting oscil-lations in the queue size, unnecessary packet delay variation and variable loss rates. The method may include a ghost queue size calculated by under-counting ghost packets for use in any RED scheme. The ghost queue size represents the length of the queue as it should have been if there was no delay.
Description
RED GHOSTS VIRTUALLY ELIMINATE
DELAYED REACTION TO PACKET LOSS
FIELD OF THE INVENTION
The invention generally relates to a method to reduce excessive packet discards due to acknowledgement delay at a queueing point in a digital conr-nrunications device such as a network node. More particularly the invention relates to maintaining a "ghost list'' of flows which have suffered a packet discard due to Active C~ueue Management (A(al~I) but: have not yet reduced their transmission rate due to acknowledgc:nrent delay. Until the transmis-sion rate is reduced, packets from ghost flows run the risk of excessive packet loss so we treat ghost packets as half visible. They rnay still be discarded but only at 1/2 the going rate. In other words the reaction by the source to a packet loss becomes virtually instantaneous. There is provision below for maintaining an augmented list of ghost. "visibility". Packets from a new ghost flow are 1/2 visible but if a packet, from this ghost flow is discarded then only 1/4 of the subsequent packets from drat flow should only be dis-carded; i.e. they are now 1/9 visible. Flow visibility would decrease with further losses and increase as retransrnissions of lost packets are received.
In addition a ghost queue size counts packets by their visibility with non-ghosts counting one. The ghost queue size estimates the length of the queue as it should have been if there was no delay and can be used in various ACM
schemes.
BACKGROUND OF THE INVENTION
In order to effect statistical multiplexing in a store and forward digital cornrnunications device, such devices will typically queue packets for subse-quent processing or transmission in a connnon storage resource such as a memory buffer. At such a queueing point, tire cornrnon storage resource may be shared by many traffic flows which cornpcae for the available bandwidth by increasing packet transmission rates. The only thing restraining the growth of transmission rates is a packet loss or tune-out. A source reacts to the lost packet or tune-out by reducing its transmission rate. Consequently tire flows will increase their transmission rate until the buffer space in bottleneck node is exhausted and arriving packets are dropped. This results in excessive packet delay.
For example, in an Internet muter, the transport level protocol of Borne sources may be some form of TCP (Transmission Control Protocol) like TCP Reno with fast recovery and fast rEaransmit (See Stevens, W., TCP
Slo~eu Sta7~t,, Congestion Avoidance, Fast Retrar~.smit, and Fast Recovery Al-gorithms, RFC 2001, January 1997.) Such soi.rrces irnplernent a window flow control which limits the number of packets from this connection allowed into the network during one Round 'I~ip Time (RTT). A source implementing TCP Reno executes congestion avoidance where the window size of each connection increases by one packet each tune a packet makes a round trip as long as no losses or tune-outs occur. Tle loss of a packet is perceived by the source when it receives duplicate acknowledgements. The source then reduces the window by half. The reduced transmission rate resulting from a packet discard is therefore perceived by the bottleneck routes only after a delay of one RTT.
To solve the problem of excessive delays at bottleneck nodes ACM schemes like RED (Random Early Detection, for further details see Floyd and Ja-cobson, Random early detection gateways ,for congestion avoidance, 1993 IEEE/AC'M Transactions on 1\etworking or, more recently, Canadian Patent Application 2291835 on December 6, 1999 1>y Ouellette, M., Aweya, J., Mon-tuno, Delfin Y., Chapman, A. entitled Load Adaptive Buffer Ma,nageme7at in Pach;et Networks ) have been proposed where an arriving packet is discarded with a probability which increases with the queue size. Incipient conges-tion is detected early causing Borne connections to reduce their transmission rate (due to a packet loss) long before buffer overflow. Unfortunately ACaM
schemes exhibit wild queue oscillations and this causes unnecessary packet delay variation and variable loss rates. The root cause of these oscillations is hysteresis due to the round trip time delay in reacting to a packet loss.
In fact the delay in reducing the window size means the chance that this flow loses another packet remains two times too high during the reaction delay period until the loss takes effect.
ACM schemes determine packet discard rates based on queue size but, due to the reaction delay, the real queue size is larger than it would have been if there was no reaction delay. This also leads to excessive packet discard.
Finally, not all flows react to packet, loss 1>y reducing their transmission rate (for example, User Datagram Protocol (UDP)). Srrch flows must be identified as misbehaving and must. be controlled by fair packet; discard (see Lin, and Morris, Dynarraics of Early Random Detectio~i., Proceedings of SIGCOM'97.)
DELAYED REACTION TO PACKET LOSS
FIELD OF THE INVENTION
The invention generally relates to a method to reduce excessive packet discards due to acknowledgement delay at a queueing point in a digital conr-nrunications device such as a network node. More particularly the invention relates to maintaining a "ghost list'' of flows which have suffered a packet discard due to Active C~ueue Management (A(al~I) but: have not yet reduced their transmission rate due to acknowledgc:nrent delay. Until the transmis-sion rate is reduced, packets from ghost flows run the risk of excessive packet loss so we treat ghost packets as half visible. They rnay still be discarded but only at 1/2 the going rate. In other words the reaction by the source to a packet loss becomes virtually instantaneous. There is provision below for maintaining an augmented list of ghost. "visibility". Packets from a new ghost flow are 1/2 visible but if a packet, from this ghost flow is discarded then only 1/4 of the subsequent packets from drat flow should only be dis-carded; i.e. they are now 1/9 visible. Flow visibility would decrease with further losses and increase as retransrnissions of lost packets are received.
In addition a ghost queue size counts packets by their visibility with non-ghosts counting one. The ghost queue size estimates the length of the queue as it should have been if there was no delay and can be used in various ACM
schemes.
BACKGROUND OF THE INVENTION
In order to effect statistical multiplexing in a store and forward digital cornrnunications device, such devices will typically queue packets for subse-quent processing or transmission in a connnon storage resource such as a memory buffer. At such a queueing point, tire cornrnon storage resource may be shared by many traffic flows which cornpcae for the available bandwidth by increasing packet transmission rates. The only thing restraining the growth of transmission rates is a packet loss or tune-out. A source reacts to the lost packet or tune-out by reducing its transmission rate. Consequently tire flows will increase their transmission rate until the buffer space in bottleneck node is exhausted and arriving packets are dropped. This results in excessive packet delay.
For example, in an Internet muter, the transport level protocol of Borne sources may be some form of TCP (Transmission Control Protocol) like TCP Reno with fast recovery and fast rEaransmit (See Stevens, W., TCP
Slo~eu Sta7~t,, Congestion Avoidance, Fast Retrar~.smit, and Fast Recovery Al-gorithms, RFC 2001, January 1997.) Such soi.rrces irnplernent a window flow control which limits the number of packets from this connection allowed into the network during one Round 'I~ip Time (RTT). A source implementing TCP Reno executes congestion avoidance where the window size of each connection increases by one packet each tune a packet makes a round trip as long as no losses or tune-outs occur. Tle loss of a packet is perceived by the source when it receives duplicate acknowledgements. The source then reduces the window by half. The reduced transmission rate resulting from a packet discard is therefore perceived by the bottleneck routes only after a delay of one RTT.
To solve the problem of excessive delays at bottleneck nodes ACM schemes like RED (Random Early Detection, for further details see Floyd and Ja-cobson, Random early detection gateways ,for congestion avoidance, 1993 IEEE/AC'M Transactions on 1\etworking or, more recently, Canadian Patent Application 2291835 on December 6, 1999 1>y Ouellette, M., Aweya, J., Mon-tuno, Delfin Y., Chapman, A. entitled Load Adaptive Buffer Ma,nageme7at in Pach;et Networks ) have been proposed where an arriving packet is discarded with a probability which increases with the queue size. Incipient conges-tion is detected early causing Borne connections to reduce their transmission rate (due to a packet loss) long before buffer overflow. Unfortunately ACaM
schemes exhibit wild queue oscillations and this causes unnecessary packet delay variation and variable loss rates. The root cause of these oscillations is hysteresis due to the round trip time delay in reacting to a packet loss.
In fact the delay in reducing the window size means the chance that this flow loses another packet remains two times too high during the reaction delay period until the loss takes effect.
ACM schemes determine packet discard rates based on queue size but, due to the reaction delay, the real queue size is larger than it would have been if there was no reaction delay. This also leads to excessive packet discard.
Finally, not all flows react to packet, loss 1>y reducing their transmission rate (for example, User Datagram Protocol (UDP)). Srrch flows must be identified as misbehaving and must. be controlled by fair packet; discard (see Lin, and Morris, Dynarraics of Early Random Detectio~i., Proceedings of SIGCOM'97.)
2 SUMMARY OF THE INVENTION
This invention reduces these oscillations by virtually eliminating the re-action delay; in effect, the buffer pretends the transmission rate is reduced immediately by labelling subsequent packets from this source as ghosts until tlue transmission rate is really reduced after one round trip time. We pro-gram the digital commrurications device to retain a'' Ghost List" of flows that have recently lost a packet. Cxhost packets picked for discard by an A(aM
scheme are only discarded at the rate that, would have resulted if there was no reaction delay. A flow remains in tl-re ghost list until the buffer actually perceives a reduction in tire transmission rate from the source. In addition this invention specifies a ghost queue size which estimates the length of the queue as it should have been if there was no reaction delay and can be used in various Active C~ueue Management (ACM) schemes. Moreover the ghost list can be used to identify misbelxaving flows.
FULL DESCRIPTION OF THE INVENTION
The digital communications device would retain a "Ghost List" of flows.
Each entry in this list contains the "flow identifier" and may contain a list of sequence ruxmbers of lost packets from this flow and/or a a "tune stamp'' and/or a "visibility factor" . A packet from a flow in tire ghost list, is called a ghost packet. Only alternate ghost, packets from the same flow picked for discard by the ACM scheme are actually discarded. If the visibility factor is implemented then actual discards would he done in a proportion consistent with the visibility factor; i.e. every fourth ghost packet picked for discard is actually discarded if the visibility factor is 1/4.
A new item is added to the ghost list when the ACM scheme discards a packet from a flow not in the ghost list. The flow identifier is stored. The sequence number of the lost packet may be stored as may a time stamp. The visibility factor, if implemented, would be set to 1/2. A flow remains in the ghost list until tire buffer actually perceives a reduction in the transmission rate from the source. This may be perceived by receipt of the packet whose sequence r-rurnber has been stored or this rnay be perceived when the time stamp expires.
The visibility factor rnay provide extra performance at the expense of extra complexity. If implemented a ghost packet picked for discard by the A(~M scheme and actually discarded causes the visibility of this flow to be
This invention reduces these oscillations by virtually eliminating the re-action delay; in effect, the buffer pretends the transmission rate is reduced immediately by labelling subsequent packets from this source as ghosts until tlue transmission rate is really reduced after one round trip time. We pro-gram the digital commrurications device to retain a'' Ghost List" of flows that have recently lost a packet. Cxhost packets picked for discard by an A(aM
scheme are only discarded at the rate that, would have resulted if there was no reaction delay. A flow remains in tl-re ghost list until the buffer actually perceives a reduction in tire transmission rate from the source. In addition this invention specifies a ghost queue size which estimates the length of the queue as it should have been if there was no reaction delay and can be used in various Active C~ueue Management (ACM) schemes. Moreover the ghost list can be used to identify misbelxaving flows.
FULL DESCRIPTION OF THE INVENTION
The digital communications device would retain a "Ghost List" of flows.
Each entry in this list contains the "flow identifier" and may contain a list of sequence ruxmbers of lost packets from this flow and/or a a "tune stamp'' and/or a "visibility factor" . A packet from a flow in tire ghost list, is called a ghost packet. Only alternate ghost, packets from the same flow picked for discard by the ACM scheme are actually discarded. If the visibility factor is implemented then actual discards would he done in a proportion consistent with the visibility factor; i.e. every fourth ghost packet picked for discard is actually discarded if the visibility factor is 1/4.
A new item is added to the ghost list when the ACM scheme discards a packet from a flow not in the ghost list. The flow identifier is stored. The sequence number of the lost packet may be stored as may a time stamp. The visibility factor, if implemented, would be set to 1/2. A flow remains in the ghost list until tire buffer actually perceives a reduction in the transmission rate from the source. This may be perceived by receipt of the packet whose sequence r-rurnber has been stored or this rnay be perceived when the time stamp expires.
The visibility factor rnay provide extra performance at the expense of extra complexity. If implemented a ghost packet picked for discard by the A(~M scheme and actually discarded causes the visibility of this flow to be
3 reduced by half. The sequence number of the lost packet is added to the list of missing sequence numbers in the ghost flow's entry in tine ghost list, together with a time stamp. When tire buffer actually perceives a reduction in tire transmission rate from the source of a ghost flow, the visibility is doubled and the associated sequence number or tune stamp are removed from the entry. If the visibility is still less than one then the flow remains in the ghost list. If the visibility is now one it is removE:d from the list.
In addition a ghost queue size counts queued packets by their visibility on arrival with non-ghosts couxrting one. The ghost, queue size estimates the length of the queue as it should have been if there was no reaction delay and can be used in various ACM schemes. For example, the RED scheme has a discard rate given by an increasing function of the queue size or average queue size. Our invention proposes drat the discard rate would be given as a function of tire ghost queue size. In this way the discard rate is a function of the queue as it should have been if there was no reaction delay.
Tlre utility of the ghost list with the visibility factor can be augmented if it is also used to detect xrrisbehaving sources. It is reasonable to declare a flow as misbehaving if an entry for this flow renra,ined in the ghost list for more than two round trip times since this would only occur if the flow failed to cut its congestion window after receiving acknowledgements that two packets had been discarded (assuming the visibility feature is implemented).
In addition a ghost queue size counts queued packets by their visibility on arrival with non-ghosts couxrting one. The ghost, queue size estimates the length of the queue as it should have been if there was no reaction delay and can be used in various ACM schemes. For example, the RED scheme has a discard rate given by an increasing function of the queue size or average queue size. Our invention proposes drat the discard rate would be given as a function of tire ghost queue size. In this way the discard rate is a function of the queue as it should have been if there was no reaction delay.
Tlre utility of the ghost list with the visibility factor can be augmented if it is also used to detect xrrisbehaving sources. It is reasonable to declare a flow as misbehaving if an entry for this flow renra,ined in the ghost list for more than two round trip times since this would only occur if the flow failed to cut its congestion window after receiving acknowledgements that two packets had been discarded (assuming the visibility feature is implemented).
4
Claims (6)
1. A method for modifying the rate of packet discard by any Active Queue Management (AQM) scheme for a buffer multiplexing multiple flows through a communications device, the method comprising:
(a) Storing a "Ghost List" of flows that have recently lost a packet due to packet discard for AQM. Packets from ghost flows are called ghost packets.
(b) Adding a new entry to the ghost list when a packet from a flow not in the ghost list is discarded for AQM. The visibility of the ghost flow is set to 1/2. The sequence number of the discarded packet is stored in the entry.
(c) Only discarding every other ghost packet chosen for discard by AQM (when the visibility is 1/2).
(d) Removing an entry from the ghost list when the buffer receives a retransmission of the discarded packet as can be verified from the packet sequence number.
(f) Calculating a ghost queue length where stored non-ghost packets count one and stored packets that are ghosts on arrival count 1/2.
The ghost queue length is used instead of the queue size for any AQM scheme.
(a) Storing a "Ghost List" of flows that have recently lost a packet due to packet discard for AQM. Packets from ghost flows are called ghost packets.
(b) Adding a new entry to the ghost list when a packet from a flow not in the ghost list is discarded for AQM. The visibility of the ghost flow is set to 1/2. The sequence number of the discarded packet is stored in the entry.
(c) Only discarding every other ghost packet chosen for discard by AQM (when the visibility is 1/2).
(d) Removing an entry from the ghost list when the buffer receives a retransmission of the discarded packet as can be verified from the packet sequence number.
(f) Calculating a ghost queue length where stored non-ghost packets count one and stored packets that are ghosts on arrival count 1/2.
The ghost queue length is used instead of the queue size for any AQM scheme.
2. The method according to claim 1, wherein an entry in the ghost list contains a timestamp set with the current time when the sequence number of the discarded packet is stored. The entry in the ghost list is removed when the timestamp expires (perhaps after one estimated round trip time).
3. The method according to claim 1, wherein an entry in the ghost list contains a visibility factor and either or both a list of sequence num-bers of discarded packets that have not yet been retransmitted and a list of timestamps giving the time when those discarded packets were discarded. This extra steps in this method comprise:
(a) Ghost packets picked for discard by the AQM scheme are actually discarded in a proportion consistent with the visibility factor; i.e.
every second ghost packet picked for discard is actually discarded if the visibility factor is 1/2 and every fourth ghost packet picked for discard is actually discarded if the visibility factor is 1/4.
(b) If a ghost packet is picked for discard by the AQM scheme and is actually discarded then the visibility of this flow would be reduced to half is former value.
(c) The sequence number of the lost packet is added to the list of missing sequence numbers in the ghost flow's entry in the ghost list together with a time stamp.
(d) When the buffer actually perceives a reduction in the transmission rate from the source of a ghost flow, the visibility is doubled and the associated sequence number or time stamp are removed from the entry.
(e) If the visibility is still less than one then the flow remains in the ghost list. If the visibility is now one the flow is removed from the list.
(a) Ghost packets picked for discard by the AQM scheme are actually discarded in a proportion consistent with the visibility factor; i.e.
every second ghost packet picked for discard is actually discarded if the visibility factor is 1/2 and every fourth ghost packet picked for discard is actually discarded if the visibility factor is 1/4.
(b) If a ghost packet is picked for discard by the AQM scheme and is actually discarded then the visibility of this flow would be reduced to half is former value.
(c) The sequence number of the lost packet is added to the list of missing sequence numbers in the ghost flow's entry in the ghost list together with a time stamp.
(d) When the buffer actually perceives a reduction in the transmission rate from the source of a ghost flow, the visibility is doubled and the associated sequence number or time stamp are removed from the entry.
(e) If the visibility is still less than one then the flow remains in the ghost list. If the visibility is now one the flow is removed from the list.
4. The method according to claim 1, 2 or 3, wherein ghost packets chosen for discard by AQM are discarded randomly with a probability equal to the visibility factor.
5. The method according to claim 1, 2, 3 or 4 wherein the ghost list is used to identify misbehaving flows as those flows that remain in the ghost list for more than two round trip times.
6. The method according to claim 1, 2, 3, 4 or 5 wherein said packet is one of: an IP packet, an AAL frame, or an ATM cell.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CA002361687A CA2361687A1 (en) | 2001-10-30 | 2001-10-30 | Red ghosts virtually eliminate delayed reactionto packet loss |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CA002361687A CA2361687A1 (en) | 2001-10-30 | 2001-10-30 | Red ghosts virtually eliminate delayed reactionto packet loss |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CA2361687A1 true CA2361687A1 (en) | 2003-04-30 |
Family
ID=4170458
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CA002361687A Abandoned CA2361687A1 (en) | 2001-10-30 | 2001-10-30 | Red ghosts virtually eliminate delayed reactionto packet loss |
Country Status (1)
| Country | Link |
|---|---|
| CA (1) | CA2361687A1 (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103188166A (en) * | 2013-03-20 | 2013-07-03 | 苏州云初网络技术有限公司 | Flow control method utilizing TCP (Transmission Control Protocol)-self-consciousness Random Early Detection (RED) |
-
2001
- 2001-10-30 CA CA002361687A patent/CA2361687A1/en not_active Abandoned
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103188166A (en) * | 2013-03-20 | 2013-07-03 | 苏州云初网络技术有限公司 | Flow control method utilizing TCP (Transmission Control Protocol)-self-consciousness Random Early Detection (RED) |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US7782758B2 (en) | Efficient loss recovery architecture for loss-decoupled TCP | |
| JP3698884B2 (en) | TCP connection performance improvement method | |
| US6535482B1 (en) | Congestion notification from router | |
| US6625118B1 (en) | Receiver based congestion control | |
| EP1251661B1 (en) | Data flow control method | |
| JP4587053B2 (en) | Communication apparatus, communication system, packet loss detection method, and packet loss detection program | |
| CN107852371B (en) | Data packet network | |
| US6980516B1 (en) | Soft, prioritized early packet discard system | |
| EP1668847B1 (en) | Encapsulating packets into a frame for a network | |
| CN111224888A (en) | Method for sending message and message forwarding equipment | |
| CA2940077C (en) | Buffer bloat control | |
| CA2361687A1 (en) | Red ghosts virtually eliminate delayed reactionto packet loss | |
| JP4838739B2 (en) | Router buffer management method and router using the management method | |
| Ayar et al. | A transparent reordering robust TCP proxy to allow per-packet load balancing in core networks | |
| CA2372023A1 (en) | Overload control method for a packet-switched network | |
| JP5046197B2 (en) | Communication apparatus, communication system, packet loss detection method, and packet loss detection program | |
| JP2003318967A (en) | Band control method, congestion control method, and network configuration apparatus | |
| Eshete et al. | On the flow fairness of aggregate queues | |
| JP5382556B2 (en) | Communication apparatus, communication system, packet loss detection method, and packet loss detection program | |
| Johny et al. | Network Dominating Attack Detection and Mitigation | |
| Fu et al. | BIPR: a new TCP variant over satellite networks | |
| Hu et al. | A block based encoding approach for improving sliding window network coding in wireless networks | |
| Chan et al. | A threshold controlled TCP for data center networks | |
| JP5170847B2 (en) | Communication apparatus, communication system, packet loss detection method, and packet loss detection program | |
| Vaidyanathan | Issues in resource allocation and design of hybrid gateways |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| FZDE | Discontinued |