CN103106153B - Based on the web cache replacement method of access density - Google Patents

Abstract

Based on the web cache replacement method of access density, the present invention relates to web cache replacement method.The present invention will solve current LRU to there is locality and LFU and there is buffer memory and pollute, the problem that hit rate is low, and provides the web cache replacement method based on access density.Method: Already in whether cache object cache pool; Whether cache pool is full; Some initialization; Delete the cache object that density value is minimum, newly-increased cache object is added cache pool; Calculate current accessed interval; Whether be back-call; Calculate access density; According to formulae discovery access density, upgrade average access interval; Upgrade correlation; Exit.The present invention is applied to internet field of storage.

Description

Based on the web cache replacement method of access density

Technical field

The present invention relates to web cache replacement method.

Background technology

Along with the development in pluralism of web data, the distribution of web page contents progressively becomes the key factor affecting Web service performance.The Data Dissemination of current main-stream adopts content distributing network technology, is redirected to the request of user from its nearest server, thus reduce access time delay and source service end load pressure.In order to effectively promote service quality, content distributing network supplier disposes content proxy server at multiple network boundary.Such as Akamai company is 1 of more than 70 countries and regions, and more than 000 network internal is affixed one's name to more than 25,000 content server.Current, content distributing network pays close attention to the routing mechanism of the selection of proxy server deployed position and content usually, but the efficiency of buffer memory is the performance key factor affecting contents distribution.Herein for the caching mechanism in web content management, propose the cache replacement policy that a kind of density mixes with content size, reduce the pressure of contents distribution flow in process and the access time delay of user with this.

Buffer memory Exchange rings depends on following two principles: 1) information of frequent access should be buffered; 2) the temperature change of information means the change of access time interval.So far existing multiple cache replacement policy is based on heavily quoting the time, current cache replacement policy mainly adopts based on frequency and local locality as replacing benchmark, but all do not consider whole access history, and there is locality and LFU and there is buffer memory and pollute, the problem that hit rate is low in current LRU.

List of references " AnovelcachereplacementpolicyforISPmergedCDN ", (the InternatioanlConferenceonParallelandDistributedSystems such as QianLi, 708-709 page, on Dec 19th, 2009), disclose a kind of new cache replacement policy of web oriented contents distribution.

Summary of the invention

The present invention will solve the low problem of buffer memory rate and byte hit that the low problem of cache hit rate that current LRU replacement method and LFU replacement method relate to and GDSF replacement method relate to, and provides the web cache replacement method based on access density.

(1) cache object whether Already in cache pool, otherwise jump into step (2), be jump into step (5);

(2) whether cache pool is full, fullly then jumps into step (3), discontented then jump into step (4);

(3) delete the cache object that access density value is minimum, newly-increased cache object is added cache pool, the access density of initialization cache object, last visit position, accessed frequency, average access interval, jump into step (10);

(4) newly-increased cache object is added cache pool, the access density of initialization cache object, last visit position, accessed frequency, average access interval, jump into step (10);

(5) Already in cache pool, calculates current accessed interval;

(6) whether is back-call, being that back-call then jumps into step (7), is not jump into step (8);

(7) make average access interval equal current accessed interval, visiting frequency+1, calculate access density, jump into step (10);

(8) not back-call, according to formulae discovery access density, upgrade average access interval;

(9) last visit position is upgraded, cache object visiting frequency+1;

(10) cache access total degree+1, exits.

Inventive principle:

One, suppose that spatial cache can hold at most M cache object, if spatial cache less than, then replacement process is consistent with other cache replacement policies;

Two, when spatial cache is expired, arrive cache object i for new, first calculate the current accessed interval now_accintvl of cache object i _i:

If current accessed interval now_accintvl _ilower than the average access interval avg_accintvl of i _i, the temperature of meaning cache object i is on a declining curve, therefore reduces the density value ad_value of cache object i _i;

If current accessed interval now_accintvl _ihigher than the average access interval avg_accintvl of i _i, the temperature of meaning cache object i is in rising trend, therefore increases the density value ad_value of cache object i _i.

Invention effect:

By analyzing actual Web data access scenarios, find that the rate of change at access interval has higher accuracy for the impact of hit rate;

First the present invention extracts the URLs (UniformResourceLocator) of real network and analyzes the access behavior of user on campus network gateway, find in the buffer memory of LRU (least recently used), the URL of high popularity is often by the replacement of some low popularities; Secondly in order to avoid this hit loss, adopt and access the weight of interval variation rate as object, and carry out buffer memory replacement in conjunction with space size shared by object;

By embodiment by this strategy respectively with LRU (least recently used), LFU (recently frequent use) and GDSF (space mixes greed with frequency) contrasts, result shows the hit rate that can promote merely 3% ~ 5% based on the replacement algorithm of accessing interval variation, and the replacement algorithm of mixing promotes the rate and byte hit of 5% ~ 8% than GDSF (space mixes greed with frequency).

The present invention adopt based on access density replacement policy CPBAD (cachepolicybasedonaccessdensity) algorithm be each cache object arrange the out-of-service time with this avoid thoroughly buffer memory pollute.Simultaneously in order to solve the storage overhead of counter, while periodic maintenance, carry out counter replacement.

Accompanying drawing explanation

Fig. 1 is module frame figure of the present invention;

Fig. 2 is the data centralization URL distribution plan in tool embodiment; A (), (b) and (c) are extracted 427 in the network log of the campus gateway of continuous three days, URL visiting frequency under the common coordinate of 936 user's requests and the graph of a relation of temperature, d (), (e) and (f) are extracted 427 in the network log of the campus gateway of continuous three days, the URL visiting frequency under the log-log coordinate of 936 user's requests and the graph of a relation of URL temperature;

Fig. 3 is the access sequence figure of the popular URL in embodiment;

Fig. 4 is that the different zipf in embodiment divide and plant the impact of λ on algorithm;

Fig. 5 is that the cache replacement algorithm hit rate in embodiment compares; A () represents the hit rate under the algorithms of different under Dataset1, b () represents the hit rate under the algorithms of different under Dataset2, c () represents the hit rate under the algorithms of different under Dataset3, d () represents the hit rate under the algorithms of different under total data collection represent LFU, represent LRU, represent CPBAD;

Fig. 6 is the rate and byte hit comparison diagram of the multiple replacement algorithm in embodiment; , represent LRU, represent LFU, represent GDSF, represent CPBADS.

Embodiment

Embodiment one: the web cache replacement method based on access density of present embodiment realizes according to the following steps:

(1) cache object whether Already in cache pool, otherwise jump into step (2), be jump into step (5);

(2) whether cache pool is full, fullly then jumps into step (3), discontented then jump into step (4);

(3) delete the cache object that access density value is minimum, newly-increased cache object is added cache pool, the access density of initialization cache object, last visit position, accessed frequency, average access interval, jump into step (10);

(4) newly-increased cache object is added cache pool, the access density of initialization cache object, last visit position, accessed frequency, average access interval, jump into step (10);

(5) Already in cache pool, calculates current accessed interval;

(6) whether is back-call, being that back-call then jumps into step (7), is not jump into step (8);

(7) make average access interval equal current accessed interval, visiting frequency+1, calculate access density, jump into step (10);

(8) not back-call, according to formulae discovery access density, upgrade average access interval;

(9) last visit position is upgraded, cache object visiting frequency+1;

(10) cache access total degree+1, exits.

Present embodiment effect:

By analyzing actual Web data access scenarios, find that the rate of change at access interval has higher accuracy for the impact of hit rate;

First present embodiment is extracted the URLs (UniformResourceLocator) of real network and is analyzed the access behavior of user on campus network gateway, find in the buffer memory of LRU (least recently used), the URL of high popularity is often by the replacement of some low popularities; Secondly in order to avoid this hit loss, adopt and access the weight of interval variation rate as object, and carry out buffer memory replacement in conjunction with space size shared by object;

By embodiment by this strategy respectively with LRU (least recently used), LFU (recently frequent use) and GDSF (space mixes greed with frequency) contrasts, result shows the hit rate that can promote merely 3% ~ 5% based on the replacement algorithm of accessing interval variation, and the replacement algorithm of mixing promotes the rate and byte hit of 5% ~ 8% than GDSF (space mixes greed with frequency).

Replacement policy CPBAD (cachepolicybasedonaccessdensity) algorithm based on access density that this embodiment adopts is that each cache object arranges the out-of-service time and avoids buffer memory thoroughly to pollute with this.Simultaneously in order to solve the storage overhead of counter, while periodic maintenance, carry out counter replacement.

Embodiment two: present embodiment and embodiment one unlike: access described in step (3) and be spaced apart this and hit cache object and hit the cache access number of times differed between this cache object last time.Other step and parameter identical with embodiment one.

Embodiment three: present embodiment and embodiment one or two unlike: access the ratio that density is the accessed number of times of cache object and the total access times of buffer memory within a period of time described in step (3).Other step and parameter identical with embodiment one or two.

Beneficial effect of the present invention is verified by following examples:

One, first the life cycle of cache object in spatial cache is analyzed;

Secondly two, analyze the behavior of web access and the distribution situation of URL: the distribution characteristics of research web request, is extracted 427 in the network log of the campus gateway of continuous three days, 936 users' requests, comprise 167,981 different URL;

Three, finally according to the temperature variation tendency of cache object, find to adopt the replacement algorithm of density to possess better predictive ability: because the locality of LRU and the buffer memory of LFU pollute character, the present embodiment proposes the variation tendency of observing cache object based on the replacement policy CPBAD (cachepolicybasedonaccessdensity) of access density within the longer period, and the weights reducing object by reducing temperature are effectively avoided buffer memory to pollute thus improved cache hit rate.

Web cache replacement method based on access density realizes according to the following steps:

(1) cache object whether Already in cache pool, otherwise jump into step (2), be jump into step (5);

(2) whether cache pool is full, fullly then jumps into step (3), discontented then jump into step (4);

(3) delete the cache object that access density value is minimum, newly-increased cache object is added cache pool, the access density of initialization cache object, last visit position, accessed frequency, average access interval, jump into step (10);

(4) newly-increased cache object is added cache pool, the access density of initialization cache object, last visit position, accessed frequency, average access interval, jump into step (10);

(5) Already in cache pool, calculates current accessed interval;

(6) whether is back-call, being that back-call then jumps into step (7), is not jump into step (8);

(7) make average access interval equal current accessed interval, visiting frequency+1, calculate access density, jump into step (10);

(8) not back-call, according to formulae discovery access density, upgrade average access interval;

(9) last visit position is upgraded, cache object visiting frequency+1;

(10) cache access total degree+1, exits.

The life cycle of the cache object in the present embodiment in step one, each cache object has self life cycle in spatial cache, and the life cycle of the object A in buffer memory can be divided into two parts: 1 active period: from entering buffer memory until accessed for the last time; The 2 ossified phases: from accessed for the last time until be replaced out buffer memory; The active period of cache object is longer, and caching performance is higher, although can not accurately predicting cache object future behaviour, but to extend its active period as far as possible or reduce the long ossified phase object time be in the buffer the effective ways promoting buffer efficiency;

Fig. 2 is the distribution situation figure of the URL in the present embodiment in step 2, (a), b () and (c) is extracted 427 in the network log of the campus gateway of continuous three days, URL visiting frequency under the common coordinate of 936 user's requests and the graph of a relation of temperature, (d), e () and (f) is extracted 427 in the network log of the campus gateway of continuous three days, URL visiting frequency under the log-log coordinate of 936 user's requests and the graph of a relation of URL temperature, blue straight line in figure is the zipf matching to red-black curve, this data set meets typical zipf _αdistribution, wherein 0.6< α <0.8, popular object means in relative short time repeatedly accessed, therefore extending these popular objects time is in the buffer the effective ways promoting buffer efficiency, and access interval is the key metrics of objective description object temperature,

Fig. 3 illustrates the access sequence of 8 URL the hottest in data centralization in step 2, y-axis represents this 8 URL, x-axis represents the order that each URL occurs in these three days, as can be known from Fig. 3, popular URL has the extremely low characteristic in interval within certain period, and the access interval of some popular URL diminishes gradually simultaneously, as url-2 and url-5, the access interval of some popular URL is comparatively even simultaneously, and as url-8, such URL generally can not be replaced in all kinds of cache replacement policy; And for this kind of URL of similar url-1, because its access interval has obvious periodicity, but period distances is longer, under the cache policy of LRU, will be replaced repeatedly, thus reduce caching performance; For this kind of URL of similar url-7, although have certain periodicity, its period distances constantly increases, and this kind of URL is under LFU strategy, and its weights rise all the time, but its temperature presents downtrending, thus causes buffer memory to pollute;

The access behavior that can obtain popular URL from above analysis can be divided into following 3 class situations:

1) evenly and have periodically, as url-8, url-4;

2) interval is accessed from sparse to dense, as url-5;

3) sudden growth, as url-3, url-6;

In view of the local locality of lru algorithm and the weights monotonicity of LFU algorithm, the present embodiment proposes the cache replacement algorithm based on access interval variation, to promote caching performance;

In order to effectively verify CPBAD algorithm,

$ad\_{\mathrm{value}}_i^n=\left\{\begin{array}{ccc}INITIAL\_VALUE,& if& {\mathrm{obj}}_{i}isanewobject\\ ad\_{\mathrm{value}}_i^{n-1}\&CenterDot;\mathrm{\&lambda;}\&CenterDot;\frac{avg\_{\mathrm{accintvl}}_i}{avg\_{\mathrm{accintvl}}_i+now\_{\mathrm{accintvl}}_i},& if& avg\_{\mathrm{accintvl}}_i<now\_{\mathrm{accintvl}}_i\\ ad\_{\mathrm{value}}_i^{n-1}\&CenterDot;\left(1+\mathrm{\&lambda;}\&CenterDot;\frac{now\_acc\mathrm{int}vl}{avg\_{\mathrm{accintvl}}_i+now\_{\mathrm{accintvl}}_i}\right),& if& avg\_{\mathrm{accintvl}}_i>now\_{\mathrm{accintvl}}_i\end{array}\right.,0<\mathrm{\&lambda;}<1---\left(1\right)$

Wherein, described C _totalrepresent the total degree that in certain period, buffer memory is accessed, last _iposition when representing that object i last time is accessed on total access sequence, now _iposition when object i last time is accessed on total access sequence, ad_value _irepresent the density value of object i, freq _irepresent total frequency that object i is accessed, avg_accintvl _irepresent the average access interval of object i, now_accintvl _irepresent the access interval that object i is current, n is the number of cache object;

Real network data set compares with LRU and LFU algorithm, and this data set is split into three sub-data sets by the date, as shown in table 1:

Table 1 gateway log data set

In order to determine the λ value in formula (1), need to determine the impact of λ on algorithm, therefore first the present embodiment generates 10,000 URL, and buffer memory is set and can holds 500 URL objects, observe algorithm and divide by changing λ the performance planted at different α, be illustrated in figure 4 different zipf and divide and plant the effect diagram of λ to algorithm represent α=0.6, represent α=0.7, represent α=0.8, represent α=0.9, represent α=1.0;

Obviously can find out that from Fig. 4 λ is when [0.6,0.8] is interval, hit rate is higher, and in experiment afterwards, the present embodiment selects λ=0.8 as empirical value;

In order to better compare replacement policy, the present embodiment is tested on 3 sub-data sets and total data collection: Fig. 5 shows the hit rate under algorithms of different, a () represents the hit rate under the algorithms of different under Dataset1, b () represents the hit rate under the algorithms of different under Dataset2, c () represents the hit rate under the algorithms of different under Dataset3, d () represents the hit rate under the algorithms of different under total data collection represent LFU, represent LRU, represent CPBAD;

Can be observed CPBAD algorithm and be better than LRU and LFU algorithm, known when cache size is 500 from Fig. 5 (c), LFU algorithm is obviously better than LRU; And known in fig. 2, the α value of data set dataset3 is greater than other 2 data sets, and it is larger that this meaning works as α value, and the hit rate of LFU is higher; Simultaneously when cache size is 2000, the hit rate of three kinds of algorithms is almost consistent, because when cache size is increased to a timing, namely the number of popular URL is close with cache size, increase spatial cache cannot promote hit rate; For Fig. 5 (d), although spatial cache rises to 8000, lru algorithm is still lower than other, and this is owing to storing the lower URL of a large amount of temperature in buffer memory; The advantage of CPBAD algorithm is can to reduce when hot data turns cold gradually its weights in spatial cache, thus promotes hit rate; When although in Fig. 5 (c), spatial cache is 2000, CPBAD and LFU is close, even when spatial cache is 500, LFU is higher than CPBAD, and this is that the buffer memory therefore in LFU keeps high hit condition always because the quantity of lasting hot data is close to spatial cache, and for CPBAD, change due to density value is weaker than LFU and keeps the freshness of buffer memory, thus some hot datas are replaced out buffer memory, and then causes hit rate to decline; But in general, CPBAD algorithm is better than LFU and LRU as a rule;

During for this replacement algorithm of actual deployment, spatial cache size can not use merely URL number as a reference but storage size, and therefore rate and byte hit is possessed of higher values for real system; Consider the diversity of web content size, the present embodiment is supposed that the size of the URL page meets between (1KB, 1MB) and is uniformly distributed; And modify for formula (1), increase file size parameter, as shown in formula (2), use weights as cache object calculate.

$size\_ad\_{\mathrm{value}}_i^n=ad\_{\mathrm{value}}_i^n+log(obj\_size/cachesize)---\left(2\right)$

Suppose that the spatial cache upper limit is that 2GB, URL distribution uses total data collection dataset in an experiment.Fig. 6 shows the rate and byte hit of often kind of cache replacement algorithm on different spatial cache, zipf α=0.645, represent LRU, represent LFU, represent GDSF, represent CPBADS, as can be seen from Fig. 6 when buffer memory is lower, GDSF (GreedyDualSizeFrequency) is better than CPBAD, but when spatial cache increases gradually, the algorithm that the present embodiment proposes obviously is better than GDSF.This is due in object weight computing, and when space is enough, some present and can not be replaced very soon in GDSF to cold large files by heat, and CPBAD can reduce ad_value value very soon, thus reduce to reach more effective space availability ratio.

Generally speaking, adopt the cache replacement algorithm of density based to have the rate and byte hit higher than GDSF for web content distribution, and then reduce the performance loss of web server.Higher rate and byte hit also means that the synchronization bandwidth consumption in a distribution Web group of planes reduces simultaneously.By test by this strategy respectively with LRU (least recently used), LFU (recently frequent use) and GDSF (space mixes greed with frequency) contrasts, result shows the hit rate that can promote merely 3% ~ 5% based on the replacement algorithm of accessing interval variation, and the replacement algorithm of mixing promotes the rate and byte hit of 5% ~ 8% than GDSF (space mixes greed with frequency).

Claims (1)

1., based on the web cache replacement method of access density, it is characterized in that the web cache replacement method based on access density realizes according to the following steps:

(1) cache object whether Already in cache pool, otherwise jump into step (2), be jump into step (5);

(2) whether cache pool is full, fullly then jumps into step (3), discontented then jump into step (4);

(3) delete the cache object that access density value is minimum, newly-increased cache object is added cache pool, the access density of initialization cache object, last visit position, accessed frequency, average access interval, jump into step (10);

(4) newly-increased cache object is added cache pool, the access density of initialization cache object, last visit position, accessed frequency, average access interval, jump into step (10);

(5) Already in cache pool, calculates current accessed interval;

(6) whether is back-call, being that back-call then jumps into step (7), is not jump into step (8);

(7) make average access interval equal current accessed interval, visiting frequency+1, calculate access density, jump into step (10);

(8) not back-call, according to formulae discovery access density, upgrade average access interval; Wherein, described according to formulae discovery access density be specially:

$ad\_{\mathrm{value}}_i^n=\left\{\begin{array}{cc}INITIAL\_VALUE,& if{\mathrm{obj}}_{i}isanewobject\\ ad\_{\mathrm{value}}_i^{n-1}\&CenterDot;\mathrm{\&lambda;}\&CenterDot;\frac{avg\_{\mathrm{accintvl}}_i}{avg\_{\mathrm{accintvl}}_i+now\_{\mathrm{accintvl}}_i},& ifavg\_{\mathrm{accintvl}}_i<now\_{\mathrm{accintvl}}_i\\ ad\_{\mathrm{value}}_i^{n-1}\&CenterDot;(1+\mathrm{\&lambda;}\&CenterDot;\frac{now\_{\mathrm{accintvl}}_i}{avg\_{\mathrm{accintvl}}_i+now\_{\mathrm{accintvl}}_i}),& ifavg\_{\mathrm{accintvl}}_i>now\_{\mathrm{accintvl}}_i\end{array}\right.,0<\mathrm{\&lambda;}<1---\left(1\right)$

Formula (1) is utilized to calculate access density according to formula (2);

$size\_ad\_{\mathrm{value}}_i^n=ad\_{\mathrm{value}}_i^n+log(obj\_size/cachesize)---\left(2\right)$

Wherein, described ad_value _irepresent the density value of object i, avg_accintvl _irepresent the average access interval of object i, now_accintvl _irepresent the access interval that object i is current, n is the number of cache object;

(9) last visit position is upgraded, cache object visiting frequency+1;

(10) cache access total degree+1, exits;

Wherein, described access is spaced apart this hit cache object and hit the cache access number of times differed between this cache object last time; Described access density is the ratio of the accessed number of times of cache object and the total access times of buffer memory within a period of time.