Toward an Understanding of the Processing Delay of Peer-to-Peer Relay Nodes
2 likes835 views
The document discusses the processing delays at peer-to-peer relay nodes and their impact on VoIP quality. A methodology for measuring these delays was proposed and validated through large-scale Internet measurement, revealing that significant processing delays can degrade VoIP performance, particularly when relay nodes are busy. It concludes that monitoring processing delays continuously is crucial for maintaining VoIP quality.
Toward an Understanding of the Processing Delay of Peer-to-Peer Relay Nodes
- 1. Understanding the Processing Delay of Peer‐to‐Peer Relay Nodes Kuan‐Ta Chen and Jing‐Kai Lou Institute of Information Science, Academia Sinica IEEE/IFIP DSN 2008
- 2. Motivation Peer‐to‐peer relaying is commonly used Purpose overlay routing (for better quality) connection establishment (for peers behind NAT/firewall) data aggregation (lower bandwidth usage) Application Conferencing: Video Streaming: Kuan‐Ta Chen / Understanding the Processing Delay of Peer‐to‐Peer Relay Nodes 2
- 3. The Problem currently relaying is achieved at application‐level (user space) Application TCP implementation scheduling delay Network Interface Card (depending on workload and scheduler) (kernel space) Kuan‐Ta Chen / Understanding the Processing Delay of Peer‐to‐Peer Relay Nodes 3
- 4. What We Do Measure the processing delays at relay nodes proposed a measurement methodology and performed a large‐scale Internet measurement (> 1000 nodes) Understand the characteristics of processing delays unreliable and predictable Examine whether these processing delays degrade application quality yes, they can significantly affect VoIP quality Kuan‐Ta Chen / Understanding the Processing Delay of Peer‐to‐Peer Relay Nodes 4
- 5. Talk Progress Overview Measuring processing delay (PD) Large‐scale measurement PD characterization Impact of PD on VoIP quality Conclusion Kuan‐Ta Chen / Understanding the Processing Delay of Peer‐to‐Peer Relay Nodes 5
- 6. Our Basic Method for Measuring PD how much delay incurred? Sender Receiver Relay node Assumptions No intentional delays at the relay node Source transmits packets with TCP TCP implementation is running at high priority, while application is running at low priority Sender and receiver are colocated at the same place Kuan‐Ta Chen / Understanding the Processing Delay of Peer‐to‐Peer Relay Nodes 6
- 7. The Idea t1 TCP data packet Sender TCP ACK PD t2 Relay node t3 Relayed packet Receiver PD = t3 ‐ t2 if network delay is approximately constant Kuan‐Ta Chen / Understanding the Processing Delay of Peer‐to‐Peer Relay Nodes 7
- 8. Verifying The Assumption: “Is TCP implemetation running at high priority?” LAN experiment Windows XP on commodity PC 200‐byte packets sent at 30pkt/sec 10 random movie clips to simulate workload X: how much time relay packets are generated Y: how much time ack packets are generated Kuan‐Ta Chen / Understanding the Processing Delay of Peer‐to‐Peer Relay Nodes 8
- 9. The Problem of Our Basic Method t1 TCP data packet Sender TCP ACK PD t2 Relay node t3 Relayed packet Receiver Network delays are NOT constant (mostly because of queueing delays) t3 – t2 != PD Kuan‐Ta Chen / Understanding the Processing Delay of Peer‐to‐Peer Relay Nodes 9
- 10. Proposed Improvement: IPID Filtering Filter out seriously delayed packets IPID: the ID field of IP packets monotonically increasing infer the order of data packets generated by the sender infer the order of ack packets generated by the relay node infer the order of relay packets generated by the relay node If any pair of packets is reordered along their transmission discarding both packets LIS (longest increasing subsequence) algorithm Kuan‐Ta Chen / Understanding the Processing Delay of Peer‐to‐Peer Relay Nodes 10
- 12. Effectiveness of IPID Filtering LAN experiments Evaluate the differences between measured PD and true PD Average error of inferred PD Maximum error of inferred PD Kuan‐Ta Chen / Understanding the Processing Delay of Peer‐to‐Peer Relay Nodes 12
- 13. Talk Progress Overview Measuring processing delay (PD) Large‐scale measurement PD characterization Impact of PD on VoIP quality Conclusion Kuan‐Ta Chen / Understanding the Processing Delay of Peer‐to‐Peer Relay Nodes 13
- 14. Large‐Scale Measurement Goal: To understand PD at real‐life relay nodes Testbed: Skype > 200 thousands super nodes available anytime Skype finds a usable relay node automatically if the caller cannot reach the calle We can force Skype to pick an alternate super node for relaying a VoIP call Kuan‐Ta Chen / Understanding the Processing Delay of Peer‐to‐Peer Relay Nodes 14
- 15. Collection Methodology Automatically call establishment and drop (10 min each) Human speech recordings form Open Speech Repository Use firewall to block caller <‐> callee direct connection We block used relay nodes to force Skype find a new one TCP packet play sound >> TCP ACK Relayed packet relay node (chosen by Skype) traffic capture Kuan‐Ta Chen / Understanding the Processing Delay of Peer‐to‐Peer Relay Nodes 15
- 17. Call Information On average, data packets incur 9 ms PD Estimated Processing Delays On average, maximum PD is around 240 ms Kuan‐Ta Chen / Understanding the Processing Delay of Peer‐to‐Peer Relay Nodes 17
- 18. Talk Progress Overview Measuring processing delay (PD) Large‐scale measurement PD characterization Impact of PD on VoIP quality Conclusion Kuan‐Ta Chen / Understanding the Processing Delay of Peer‐to‐Peer Relay Nodes 18
- 20. A Classification of PD Patterns (1) Typical Processing Delays variation of PDs is small magnitude is rarely higher than 20 ms the relay node is lightly‐loaded and the computer is not in use Kuan‐Ta Chen / Understanding the Processing Delay of Peer‐to‐Peer Relay Nodes 20
- 21. A Classification of PD Patterns (2) Varible Processing Delays PDs are stable most of time, but occasionally very different e.g., the PDs of call 561 are relatively high during the 250–450 ms period, which is likely due to a person using the computer, or an application with a time‐varying workload is running. Kuan‐Ta Chen / Understanding the Processing Delay of Peer‐to‐Peer Relay Nodes 21
- 22. A Classification of PD Patterns (3) Level‐shifted Processing Delays The levels of PDs are increased or decreased by a significant magnitude, say, larger than 10 ms likely a heavily loaded task starts or stops running on the relay node Kuan‐Ta Chen / Understanding the Processing Delay of Peer‐to‐Peer Relay Nodes 22
- 23. A Classification of PD Patterns (4) Periodic Processing Delays Bursts of high PDs occur at regular intervals, possibly because of the behavior of an application. e.g., the 1‐minute interval in Call 214 may be caused by an email notification program with a one‐minute check interval. Kuan‐Ta Chen / Understanding the Processing Delay of Peer‐to‐Peer Relay Nodes 23
- 24. A Classification of PD Patterns (5) Loaded Processing Delays The level of PDs remains large, say 100 ms or higher. the relay node should be running computation‐ or I/O‐ intensive applications Kuan‐Ta Chen / Understanding the Processing Delay of Peer‐to‐Peer Relay Nodes 24
- 25. Busy Levels The level of workload on the relay node 95% percentile of PD within a 10‐second window Lightly‐load host: application always serviced when needed Even on a heavily load machine, the PD incurs by relay packets are not always large A packet arrives just before the application’s execution quantum get serviced immediately otherwise has to wait some time If busy level changes > 10 ms in successive 10‐second windows unstable relay nodes Kuan‐Ta Chen / Understanding the Processing Delay of Peer‐to‐Peer Relay Nodes 25
- 26. Average Ratios of All Relay Nodes and Unstable Relay nodes Unstable nodes are mostly observed in work hours at local time Kuan‐Ta Chen / Understanding the Processing Delay of Peer‐to‐Peer Relay Nodes 26
- 27. Talk Progress Overview Measuring processing delay (PD) Large‐scale measurement PD characterization Impact of PD on VoIP quality Conclusion Kuan‐Ta Chen / Understanding the Processing Delay of Peer‐to‐Peer Relay Nodes 27
- 28. Impact of PD on VoIP Quality p2p relaying is commonly used by VoIP applications bypass firewall/NAT network latency may be reduced by overlay routing VoIP bandwidth requirement is not high, e.g., 32 Kbps Use trace‐driven simulations to assess the impact of PD combine different network delays and processing delays simulate VoIP calls with and without processing delays examine their differences in VoIP quality Kuan‐Ta Chen / Understanding the Processing Delay of Peer‐to‐Peer Relay Nodes 28
- 29. Methodology Measuring VoIP quality degradation ITU‐T E‐model (G.711) based on delay and loss rate yields a R‐score ranges from 0 ‐‐ 100 R‐score > 80 satisfactory R‐score < 70 unacceptable Two classes of VoIP playout buffer sizing schemes static buffer (60 ms according to Skype) adaptive buffer (according to Ramjee INFOCOM’94) size = mean delay + 4 x standard deviation of delays mean and std dev are derived from EWMA processes 50,000 random samples from 1,243,225 (1,115 x 1,115) possible configurations Kuan‐Ta Chen / Understanding the Processing Delay of Peer‐to‐Peer Relay Nodes 29
- 30. R‐scores Decrease due to PD 20% of calls incur significant 40% and 60% of calls incur significant average R‐score decrease maximum R‐score decrease PD leads to larger playout buffer that leads to lower e2e packet loss rate Kuan‐Ta Chen / Understanding the Processing Delay of Peer‐to‐Peer Relay Nodes 30
- 31. Original R‐scores vs. Average R‐scores more degradation than static buffer with original scores >= 40 1. higher scores => lower net delay 2. PD impact buffer size more Kuan‐Ta Chen / Understanding the Processing Delay of Peer‐to‐Peer Relay Nodes 31
- 32. Impact of Busy Levels What busy levels of PD would significantly impact VoIP quality? busy level > 20 ms significant quality degradation Kuan‐Ta Chen / Understanding the Processing Delay of Peer‐to‐Peer Relay Nodes 32
- 33. Busy Period Characterization We define a node as busy when its busy level is higher than 20 ms 23% of relay nodes were ever busy during a 10‐minute call mean: 18 sec mean: 25 sec 65% shorter than 10 sec The nodes tend to switch between busy and non‐busy states frequently Kuan‐Ta Chen / Understanding the Processing Delay of Peer‐to‐Peer Relay Nodes 33
- 34. Summary consider a hidden aspect of peer‐to‐peer relaying ‐‐‐ processing delays at relay nodes PD can be determental to VoIP quality we should avoid a relay node with a busy level higher than 20 ms we have to monitor the processing delays of a relay node continuously (like we do for the network delays) Kuan‐Ta Chen / Understanding the Processing Delay of Peer‐to‐Peer Relay Nodes 34
- 36. Thank You! Kuan‐Ta Chen (Chun‐Yang Chen) http://www.iis.sinica.edu.tw/~ktchen IEEE/IFIP DSN 2008