Transcript ppt
ON THE STABILITY OF SKYPE SUPER NODES Anat Bremler-Barr Ran Goldschmidt Interdisciplinary Center Herzliya [email protected] Haifa University [email protected] Background: • VoIP • P2P Technology • Proprietary signaling and media protocol Architecture: • Based on partially centralized P2P networks • Two types of peer nodes: • Regular Clients and Super Nodes • Super Nodes (SNs): • Control level – heart of Skype • Super Nodes = Skype Clients with good Bandwidth, CPU usage ... The Role of a Super Node (SN) • Maintains control information: the IPs of the Skype users • Each client maintains an SN list subset of SNs • SN list is constantly updated Call Bob • Client that wishes to use Skype (to call) picks one SN from the SN list • Querying the IP of the callee • SN is defined by (IP,Port) IP=12.3.2.4, Port=3 Main goal: Understanding the stability of SNs • P2P networks have an inherent dynamic nature • Clients and computers join and leave the network constantly • Choosing stable SNs is an important task: • Improve the performance and quality of the P2P network • Our goal: measure the stability of SNs and understand how to choose stable SNs Experiment Methodology Stage 1: Collecting 10,000 SNs in 15 minutes Stage 2: Ping each SN every 15 minutes for 3 months UDP packet of Skype login Harvesting of SNs: Using the SN list - version 2.5 holds a list of 200 SNs Absolute Availability • Absolute Availability = the percentage of time the SN is up during the test 18% -16 days • Low: 50% of the SNs are available less than 18% of the test time, < 16 days Number of Sessions of SN • A session is the continuous period of time the node is up. • 40% of the SNs have one session! MSLT vs SSLT We define two group of SNs: o o Single Session in Life Time (SSLT) Multiple Sessions in Life Time (MSLT) We analyze the different characteristics of the two groups We play “what if” game and show the stability of the system if we take SNs only from one group (SSLT or MSLT) Residual life Residual life time - time elapsed between the harvesting of the SN until the end of its last session 1.75 67.5 • SSLT (one session) : median of residual life is 1.75 days • MSLT (multi session) : median of residual life is 67.5 days First Session Length • SSLT (one session) : median of session is 1.75 • MSLT (multi session) : median of session is 4.35 Surprise ! Super Nodes in SSLT have only one session and a shorter one… The impact on the System: Churn • Churn measures the number of times an SN goes down and we need to replace the SN. • Assumption: system maintains a fixed number of SNs • When an SN fails, the system picks another one to replace it. • When an SN fails and then recovers, the SN is like a new SN. Result: SSLT: 0.35 turnovers per day MSLT: 0.22 turnovers per day The impact on the System: Accessibility • Accessibility: the probability that an SN is alive as a function of time • If an SN fails and then recovers, the SN is viewed as a node that never failed. • Impact: The lower accessibility value the higher the update rate of the SN list Skype Accessibility • The SNs are distributed over the world good accessibility. • The SNs in the SN list are also distributed over the world. 50 Percentage of SNs(10,000) 45 40 35 30 25 20 15 10 5 0 Others Asia Europe North America SSLT and MSLT Summery • A huge difference between the characteristics of SSLT and MSLT SNs • Choosing SNs from the MSLT group would improve the churn and accessibility of the P2P system dramatically Why SSLT and MSLT are so different ? Our Answer: Static vs. Dynamic IPs • SSLT = dynamic IPs • Residential users: Cable, xDSL… • The address is changed from time to time (or between sessions) • SN died since the IP address of the SN was replaced • Good chance that this SN is alive but with a different IP address • MSLT = static IPs • SN can leave and return with the same IP • Servers or academic networks high availability good infrastructure longer session Classifying SNs according to the address type Classifying using DNS information • The vast majority 84.92% of the static IP SNs belongs to the MSLT group • The dynamic IPs group shows weaker correlation with only 61.45% of the SNs belongs to SSLT group • We think that this is due to “Sticky Dynamic IPs” – DHCP technology • Client that is disconnected from the network returns to its old IP if the break is short Related Work • • Our work: Skype Super Nodes Most previous work: on file transfer Check: correlation between ICMP and Skype Ping Skype: application is on all the time the computer is on File transfer: application is on only when performing file transfer File Transfer Skype Session Short (hours) Long (days) Dynamic Application impact Computer impact (IP address type effect) Summary & Practical Implication Practical implication: choosing SNs with static IPs stable P2P network Note: Classifying the IP is an easy task for the P2P application. The high stability of static IPs is due to two reasons: 1. Static IPs do not change the IP address - Impact the Accessibility 2. Computers with static IPs are more stable computers - Impact the Churn