220 likes | 317 Views
Wireless Cache Invalidation Schemes with Link Adaptation and Downlink Traffic. Presented by Ying Jin. Outline. Background System model Three proposed strategies Simulation Result Conclusion. Background. Cache invalidation strategy -- IR (invalidation report)
E N D
Wireless Cache Invalidation Schemes with Link Adaptation and Downlink Traffic Presented by Ying Jin
Outline • Background • System model • Three proposed strategies • Simulation Result • Conclusion
Background • Cache invalidation strategy -- IR (invalidation report) • Server periodically broadcast IR, IR ={(Ti,<dx,tx>|tx>Ti - ωL} • If cache miss, client send uplink request to server • Server collect all requests and broadcast replies once every IR period • To answer a particular query, a client is required to wait for the next IR to determine whether its cache is valid or not. • Advantages: • high scalability • energy efficiency • Drawbacks: • clients must flush their entire caches after long disconnection (ωL), even if some of the cached items may still be valid; • clients must at least wait for the next IR before answering a query to ensure consistency.
Background • Cache invalidation strategy -- IR (invalidation report) • To answer a particular query, a client is required to wait for the next IR to determine whether its cache is valid or not.
Background • IR-UIR : • UIR ={(Ti,<dx,tx>|tx>Ti}, updated IR • Client use UIRs to invalidate cache data • Reduce the long query delay • Little more broadcast overhead
Background • IR+UIR : • UIR ={(Ti,<dx,tx>|tx>Ti} • Reduce the long query delay • Little more broadcast overhead • Assumptions: • broadcast channel is error-free, • no other downlink traffic. • Objective • Study performance of IR, IR-UIR on realistic system model • Effect of broadcast overhead on other downlink traffic • Three new schemes
Some concepts • Fast fading: fluctuating in a very fast manner (caused by multi-path signals interfering with each other) • Long-term fading: fluctuating in relatively slower manner (due to distance and terrain effects) • Coherence time: time duration of the radiation maintains a near-constant phase relationship • Channel State Information (CSI) : channel condition (fading attenuation)
System model • Downlink: • Acknowledgement • Polling • Information • Announcement • Uplink: • Request • Information • Pilot Frame duration: 2.5ms
System model • System model with an adaptive physical layer • Two signal propagation components: • fast fading component • long-term shadowing component • Transmission mode • mode 0 to mode 5 (Low rate to high rate) • Assumption: • mobility of the users < 5km/hr (pedestrian speed) • channel fading experienced by each mobile device is independent of one another.
Proposed methods • Targets: • reducing the probability of corruption in IRs, • improving the broadcast channel utilization, • reducing the average delay in other downlink traffic. • Notation • User: • Voice: rspeech Kbps • Data: rfile Kbps, exponentially distributed request mean arrival time Tq • Tu: mean data update time, exponential distribution • Pu: probability of updating hot data iterm • Each server has consistent view of DB, broadcast same set of IR+UIR • Broadcast scheduler: determine transmission rate for broadcast
Proposed methods 1 • Reducing the Probability of Corruption in IR • Time interval: L seconds • # of UIR: m-1 • IR => {IRi, i= 1,2,… ω}, • IRi = {(dx,tx)| Ti- j*L < tx≤(Ti-(j-1)*L} • each segment IRi separately transmitted • For example, IR at Ti <= IR at Ti-3, IR1, IR2 at Ti • Reduce both the corruption probability and power consumption • (1-Pe)(SωL +x) < (1-Pe)(SL+x) , (Pe bit error rate, SL size of an IR segment, SωLsize of IR)
Proposed methods 2 • Improving Channel Utilization • Optimal transmission rate • current channel status of all clients • importance of the information being delivered • more important information: low-rate broadcast (higher level of error protection) • less important information: high-rate broadcast (lower level of error protection) • Two type users: • Active user: latest IR segments • long time disconnected user: old IR segments • Broadcast scheduler: • using average data rate (by collecting CSIs)
Proposed methods 3 • Reducing the Average Delay in Other Downlink Traffic • IR based scheme => block other downlink traffic • Server collect all requests over the IR time period, and broadcast after IR • size of each IR is very large • long list of reply • Server broadcasts query replies after both IRs and UIRs • Reduce block in other downlink traffic • Reduce query delay • Tradeoff between aggregate effect
Simulation results • Model • Parameters • Transmission mode • 0-5 low-> high • Three Metrics • Average query delay • # of uplink request per successful query • Average delay of other downlink traffic
Simulation results • Effect of number of clients • # of client increase => query delay decrease • IR-UIR worse than IR on aggressive broadcast • Divide-IR outperforms significantly on aggressive broadcast • UIR-reply on normal broadcast better than ideal IR-UIR • More cache hits => decrease uplink request • IR better than IR+UIR in uplink request? • Conservative broadcast achieves the least transmission error, its impact on other traffic is largest. (because long broadcast time ) Tu= 100s; Tq = 100s
Simulation results • Effect of Query Generation Time • Tq increase => query delay increase • Divide-IR not very effective • UIR-IR perform better • Tq increase => uplink request increase • Ideal IR request fewer uplink request • Tq increase => delay in other downlink traffic decrease # of client= 50; Tu= 100s
Simulation results • Effect of Update Arrival Time • Larger Tu => small delay • Divide-IR improve significantly for aggressive broadcast • UIR-reply outperform Divide-IR at high update rate • Uplink request decreases with increasing update time # of client= 50; Tq = 100s
Simulation results • Effect of Number of UIR • More UIR =>smaller delay, larger overhead • Optimal UIR=5 • Divide-IR improves with UIR • Uplink cost start to converge from UIR=5 • UIR overheads => increase delay in other downlink traffic # of client= 50; Tu= 100s; Tq = 100s
Simulation results • Effect of Access Skew • Hot data access probability • Divide-IR shows large improvement on aggressive broadcast • Cache hit => Access skew largely affect uplink cost • Delay in other downlink is comparatively not affected? # of client= 50; Tu= 100s; Tq = 100s
Simulation Results • Effect of Disconnection Time • Short Disconnection period, no big difference (fig. a) • Flush the whole cache => Significant increase in the number of uplink request (fig. d) • Little improvement in UIR-reply (fig. b) • Long disconnection => decrease query rate (fig. e) # of client= 50; Tu= 100s; Tq = 100s
Conclusion • Assumptions on IR-based cache invalidation strategies • Error-free broadcast • No other downlink traffic • Three new schemes • Divide-IR • Adaptive broadcast transmission • UIR-reply • Simulation result • Contributions • Estimate the performance of IR, IR-UIR on a realistic environment • Take into account the transmission error and other downlink traffic