A Strategy to Manage Cache Coherency in a Distributed Mobile Wireless Environment

Transcription

1 A Strategy to Manage Cache Coherency in a Distributed Mobile Wireless Environment Anurag Kahol, Sumit Khurana, Sandeep K. S. Gupta, and Pradip K. Srimani Shumei Chen Hoang Nguyen Jeff Wilson Team Members 11/14/01 2 1

2 Outline Introduction Proposed Scheme & Comparison with Existing Schemes Formal Description of the Proposed Scheme Performance Analysis Performance Comparison Conclusions 11/14/01 3 Objectives Introduction Minimize overhead for MH to validate cache upon reconnection Allow stateful servers Minimize bandwidth requirement 11/14/01 4 2

3 Approach Introduction (cont.) Use asynchronous cache invalidation messages Buffer invalidation messages from servers at MH s Home Location Cache (HLC) when MH is disconnected, and deliver upon reconnection 11/14/01 5 Results Introduction (cont.) Average data access latency is reduced Numbers of up-link requests by MH decrease Cost is buffers (and buffer management) at HLC Validation Mathematical model Simulation 11/14/01 6 3

4 Background Caching improves performance in distributed systems Query response time Throughput Caching concerns What to cache Invalidation Consistency 11/14/01 7 Background (cont.) Mobile computing issues Low power (battery) Frequent disconnects (planned and unplanned) Mobility 11/14/01 8 4

5 Background (cont.) Cache maintenance is intended to maintain consistency must conserve battery power Energy efficient Adapt to varying QoS 11/14/01 9 Background (cont.) Caching scheme factors Data access patterns and update rates Communication/access costs Client mobility pattern Connectivity characteristics (disconnection pattern, available bandwidth,.) Location dependence of the data 11/14/

6 Background (cont.) Cache invalidation Wired approach (e.g., NFS) not suitable for wireless Most mobile environment schemes are based on a call-back (invalidation report) mechanism Must account for cases where MH may be disconnected at time invalidation report is sent 11/14/01 11 System model 11/14/01 12 A. Kahol, S. Khurana, S. K. S. Gupta, P. K. Srimani, An Efficient Cache Maintenance Scheme for Mobile Environment 6

7 States of Operation Data Hoarding Reintegration Disconnected 11/14/01 13 Mobile Host States Wake-up Awake (Active, Dozing) Sleep 11/14/

8 Communication Up-link (query) Down-link (reply) 11/14/01 15 Caching MH caches portion of database in memory to reduce number of up-link requests (queries) required The challenge is to ensure client cache remains consistent with shared database 11/14/

9 Assumptions Whenever any data item anywhere is updated, an invalidation message is sent to all MSS via wired network (perfect connectivity) MH can detect network connection state MH informs HLC before it stores (or updates) any data item in its local cache (implies connection is required) 11/14/01 17 Assumptions (cont.) Static host which is nearest to the MH, and maintains HLC of the MH, forwards to the MH any invalidation reports it receives 11/14/

10 Proposed Scheme Pre-assigned static MSS maintains MH home location cache (HLC) For a roaming MH, its HLC is duplicated at the MSS of its current cell An MSS always maintains an HLC for each MH in its coverage area at any given time 11/14/01 19 Proposed scheme (cont.) Each HLC being maintained in the MSS keeps track of what data have been locally cached at MH The HLC is a list of records (x,t,invalid_flag) for each data item x locally cached at MH (x is the ID of data item and T is the timestamp of the last invalidation of x) 11/14/

11 Proposed scheme (cont.) Invalidation reports are transmitted asynchronously and are buffered at the MSS until explicit acknowledgment is received from the specific MH. The invalid_flag is set to true for data items for which an invalidation report has been sent to the MH but no acknowledgment is received 11/14/01 21 Proposed scheme (cont.) Each MH maintains a local cache of data items it frequently accesses. Before answering queries from application, it will verify whether requested data is in consistent state. If data is valid, it will satisfy the query, otherwise an uplink request to MSS is needed 11/14/

12 Proposed scheme (cont.) When MH receives invalidation report from MSS, it marks a particular data item in its local cache invalid Cache time stamp is used to allow MA to discard invalidations it no longer needs to keep and determine which invalidation it needs to resend to MH 11/14/01 23 Proposed scheme (cont.) MA discards any messages from MH s HLC with time stamp less than or equal to timestamp t it received from MH and only sends invalidation messages with timestamp greater than t 11/14/

13 11/14/01 25 A. Kahol, S. Khurana, S. K. S. Gupta, P. K. Srimani, An Efficient Cache Maintenance Scheme for Mobile Environment Scheme comparison All existing schemes are based on synchronous or periodic broadcasting of invalidation reports All existing scheme are based on aggregating queries for a fixed period of time and then answering them after receiving an invalidation report 11/14/

14 Scheme comparison (cont.) All existing schemes assume a stateless server and do not address mobility issues Under existing schemes, the entire cache is invalid if the client is disconnected for a period larger than broadcast period or some multiple of it 11/14/01 27 Scheme comparison (cont.) Poor network utilization under existing schemes (no traffic for a long period of time followed by heavy burst) Higher queuing delay under existing schemes Proposed scheme is asynchronous and stateful 11/14/

15 Scheme comparison (cont.) Performance is enhanced at cost of maintaining additional buffer spaces at MSS 11/14/ /14/01 30 A. Kahol, S. Khurana, S. K. S. Gupta, P. K. Srimani, An Efficient Cache Maintenance Scheme for Mobile Environment 15

16 11/14/01 31 A. Kahol, S. Khurana, S. K. S. Gupta, P. K. Srimani, An Efficient Cache Maintenance Scheme for Mobile Environment Formal Description Data Structure Each data object has an unique identifier Data x denotes data associated with a data item with identifier x 11/14/

17 MH data structure t s : Time stamp of last invalidation report or data received by MH from its home MSS cache: Data cache. Data cache item is of the form (x, Data x, Valid_flag). Data x is considered valid only if Valid_flag is TRUE. 11/14/01 33 MH data structure (cont.) First_Request : A flag set to TRUE when an MH wakes up and is yet to make its first query after awakening. It is set to FALSE once the first request after waking up is made. First_Waiting : A flag set to TRUE when an MH has made its first query after getting reconnected to the network but the data has not been received. 11/14/

18 MSS data structure An array of list HCL[1.N] where HCL[i] is a list of record of type (x,t,invalid_flag) for each data item x cached by Mh i. N is the total number of MHs that are in the MSS cell. T is the time stamp of the last invalidation of x. The invalid_flag is set to TRUE for data items for which an invalidation report has been sent but no implicit acknowledgment has been received. 11/14/01 35 Messages INVALIDATION_REPORT (item_list,t,first_flag): Sent to MH by its home MSS to report invalidation of data in item_list. T is the time stamp associated with this report. first_flag is set to TRUE if this invalidation report is in response to the first query of the MH upon waking up. 11/14/

19 Messages (cont.) DATA_REQUEST(i,x,t,first_request): Sent by MH i to its MSS to request data item x when x is not found in its local cache. t is set to the time stamp t s maintained by the MH. first_request is set to TRUE if this is the first data request after the MH regains connectivity to the network. DATA(x,Data x,t): Broadcast by an MSS to send data to all MHs caching x. T is a time stamp set to the current time at the MSS. 11/14/01 37 MSS tasks If MSS receives a request for data from an MH (DATA_REQUEST): For each request, an MH sends the time stamp of the last message it had received from the MSS. The MSS deletes all entries in the HLC for the MH which had been invalidated before the time stamp carried in the message. If the data request is the first after a sleep, all items cached by MH, marked invalid since the last message received by MH are repeated through an invalidation report. Requested data item is sent to MH and added to its HLC. 11/14/

20 MSS tasks (cont.) Data item(s) updated at MSS (informed by server): The MSS sends an invalidation report to all MHs that are caching the changed data item. It also updates the data time stamp in the HLC for these MHs and marks these items as invalid. 11/14/01 39 MH tasks MH generates a request for a data item: If the MH has woken up after a sleep and if it is a first request, it sends a data request for the item and sets a flag first_waiting to TRUE to indicate that the buffered invalidations during the sleep period have not yet been received. Once the invalidation reports are received, it will answer successive queries from the cache. If the query is not a first after waking up, the cache is checked for that data item. If the item is not in the cache, the data request is sent to the MSS and the query is answered once the data arrive from the MSS. 11/14/

21 MH tasks (cont.) MH receives an invalidation report from the MSS: The MH set its cache time stamp to the time stamp in the current message and invalidates all the data items mentioned in the report in its cache. 11/14/01 41 MH tasks (cont.) MH receives data from the MSS: MH updates its cache with the current information and also updates its cache time stamp. MH wakes up after disconnection: MH set the flag First_Request to TRUE 11/14/

22 Performance Analysis Performance Analysis Goal: estimate miss probability and mean query delay Model constraints: single Mobile Switching Station (MSS) with N mobile hosts 11/14/

23 Assumptions M data items, each of size b a bits No queuing of queries when MH is disconnected Single wireless channel of bandwidth W All messages are queued and serviced FCFS 11/14/01 45 Assumptions (cont.) Queries are of size b q bits; invalidations are of size b i bits Processing overhead is ignored 11/14/

24 MH queries Queries have Poisson distribution with mean rate? Queries are uniformly distributed over all items M in the database P( N( t) = n) = n ( λt) λt n! e 11/14/01 47 MH queries (cont.) Time between updates follows exponential distribution with mean 1/µ Update rate is exponentially distributed with mean µ 11/14/

25 11/14/01 49 A. Kahol, S. Khurana, S. K. S. Gupta, P. K. Srimani, An Efficient Cache Maintenance Scheme for Mobile Environment MH states Mobile hosts alternate between sleep and awake modes s is the fraction of time in sleep mode; 0 <= s <= 1? is frequency at which state changes Consider exponentially distributed interval of time t with mean 1/? 11/14/

26 11/14/01 51 Hit ratio estimation? e = (1 s)? = effective rate of query generation Queries are uniformly distributed over M items in the database? x =? e /M = [(1 s)?/m] 11/14/

27 Hit rate estimation (cont.) Queries for a specific data item x by a MH would be a miss in the local cache (which would require an uplink request) in either of two conditions [Event 1] During time t, item x has been invalidated at least once [Event 2] Data item x has not been invalidated during time t, but MH has slept at least once during t 11/14/ /14/01 54 A. Kahol, S. Khurana, S. K. S. Gupta, P. K. Srimani, An Efficient Cache Maintenance Scheme for Mobile Environment 27

28 t λ xt ( λxe ) µ x µ Mµ P[Event 1] = µ e dx dt = = 0 0 λx + µ (1 s) λ + Mµ Probability of query for x Probability of invalidation in time [0,t] 11/14/01 55 Probability of state change P[Event 2] = t λxt µ t λxt1 ωt1 ω ( t t1 ) ( λ e e ) e ωe ( 1 e ) x 0 0 dt dt 1 Probability of query for x Probability of at least one wakeup in [t-t 1 ] Probability of invalidation in time [t,8 ] Probability that query for x is after t 1 11/14/

29 29 11/14/01 57 ( ) ( ) ( ) = [Event 2] t t t t t t t x dt dt e e e e e P x x ω ω λ µ λ ω λ ( ) = ω λ µ ω ω λ µ λ ω λ µ λ λ ω λ ωλ x x e x e e e x 2 11/14/01 58 P miss and P hit P miss = P[Event 1] + P[Event 2] P hit = 1 - P miss

30 Mean query delay estimation T delay = mean query delay Model up-link queries as M/D/1 queue (assumes dedicated up-link channel of bandwidth C) Model invalidations on down-link channel as M/D/1 queue (assumes dedicated down-link channel of bandwidth C) 11/14/01 59 N MHs in cell Uplink query generation rate Query service rate 11/14/01 60 A. Kahol, S. Khurana, S. K. S. Gupta, P. K. Srimani, An Efficient Cache Maintenance Scheme for Mobile Environment 30

31 Mean invalidation arrival rate Invalidation service rate 11/14/01 61 A. Kahol, S. Khurana, S. K. S. Gupta, P. K. Srimani, An Efficient Cache Maintenance Scheme for Mobile Environment Effective arrival rate of invalidations Query service rate 11/14/01 62 A. Kahol, S. Khurana, S. K. S. Gupta, P. K. Srimani, An Efficient Cache Maintenance Scheme for Mobile Environment 31

32 Mean query delay estimation Combine M/D/1 queues Resultant mean query delay T delay = P miss T q T q ~ 2µ q λq + λi = ~ 2µ q µ q λq + λi Average delay by an uplink query 11/14/01 63 Performance Comparison Simulation experiment Model Single cell with a base station and varying number of mobile hosts High/low invalidation rate Goal Experimental results vs. theoretical analysis Efficiency in managing disconnection Metrics: Uplink,Delay 11/14/

33 Performance Comparison Performance Comparision Low Data Invalidation Rate 11/14/01 66 A. Kahol, S. Khurana, S. K. S. Gupta, P. K. Srimani, An Efficient Cache Maintenance Scheme for Mobile Environment 33

34 Performance Comparison High Data Invalidation Rate AT > TS 11/14/01 67 A. Kahol, S. Khurana, S. K. S. Gupta, P. K. Srimani, An Efficient Cache Maintenance Scheme for Mobile Environment Impact factors Network delay and waiting for broadcast delay TS > AT Network congestion, total query Cache discard Size of invalidation report 11/14/

35 Comparison with Ideal scheme 11/14/01 69 A. Kahol, S. Khurana, S. K. S. Gupta, P. K. Srimani, An Efficient Cache Maintenance Scheme for Mobile Environment Model Validation 11/14/01 70 A. Kahol, S. Khurana, S. K. S. Gupta, P. K. Srimani, An Efficient Cache Maintenance Scheme for Mobile Environment 35

36 Pros of AS scheme Deal with frequent disconnection Minimize overhead Preserve bandwidth Reducing latency 11/14/01 71 State information management State information about local cache maintained at the home MSS Keep consistency with local cache Coda s state info scheme Info stored at server part and MSS part 11/14/

37 Data concurrency Cache discard after disconnection Coda s disconnection handling AS targets applications asking for strict data currency guarantee and where stale data access is undesirable Bank account, traffic information.. 11/14/01 73 Application in mobile IP implementation Mobile IP architecture Home agent Foreign agent Integrated with AS MA s located in MSS HLC maintained in home agent, can be duplicated on foreign agent 11/14/

38 11/14/01 75 A. Kahol, S. Khurana, S. K. S. Gupta, P. K. Srimani, An Efficient Cache Maintenance Scheme for Mobile Environment Summary Bandwidth saving and latency reduction proved through performance analysis and simulation Future work includes study the effect of different mobility pattern, cost of invalidation 11/14/