Academia.eduAcademia.edu

Figure 3 – uploaded by Phan Quang Vinh

See full PDFdownloadDownload figure
the first (P}), second (P) and third regions (Nj), i.e., four

Table 3 the first (P}), second (P) and third regions (Nj), i.e., four

Related Figures (11)

Table 1 Summarized characteristics of replacement strategies Towards the above issues, this paper presents a so-called Learning Based Replacement Algorithm (LBR), a new type of web cache management policy by applying a machine learning model to learn usage patterns from past history for web cache replacement policy, and exploiting it to select objects for replacement [5, 10, 14-16, 26, 27, 30, 31]. The features considered include frequency of reference, size Frequently Used(LFU) ranks objects based on frequency of reference. Thus, the least referred objects come on the top of the replacement of reference history resulting in wasting recency (temporal ist. LFU’s weakness in partial aspect may trigger keeping objects too long, cache space. While LRU is ocality), LFU depends on frequency (popularity). To reduce cache load, some studies based on proposed an aging mechanism that will evict long-lived objects, such as LFUDA [6], GD 6], GDS [6] and GDSF [1, 7, studies showed that size- hit rate while frequency- based meth based methods obtained 8], Some a better a better ods achieved byte hit rate. Some methods occupied more complicated techniques using statistical approach for data hooding. The typical one is the Lowest posed by Lorenzetti et al. model to calculate the re Relative Va [20]. The L ative value ue (LRV) policy, pro- RV uses a cost/benefit of each document in the cache, i.e., (CXP)/B, where C is the size of the requested document, expressing the overhead of ument from the original site, B is t transferring the doc- he bandwidth toward the server, and P is the probability that a document will com ful LWU is regarded as d of t to both recency and Anyway, the drawbac of cached documents. than LFU, LRU and other size-based be accessed again. In general, the LRV performs better k of As another we Least Unified-Value ( Usage A a model-driven simulation. The ynamic and sel mance seems better than ot techniques, but with plicated mechanism. However, it is particularly use- for small cache. For a larger cache, its performance drops. The Least Weighted icy was proposed in [23] as gorithm (LWU) pol- f-adaptive. Its perfor- her caching algorithms in terms he hit-rate. LWU can perform well with its adaptation frequency in di LWU is that it ignores the size fferent environments. 1-known method, the LUV) policy was proposed by Bahn
Table 1 Summarized characteristics of replacement strategies Towards the above issues, this paper presents a so-called Learning Based Replacement Algorithm (LBR), a new type of web cache management policy by applying a machine learning model to learn usage patterns from past history for web cache replacement policy, and exploiting it to select objects for replacement [5, 10, 14-16, 26, 27, 30, 31]. The features considered include frequency of reference, size Frequently Used(LFU) ranks objects based on frequency of reference. Thus, the least referred objects come on the top of the replacement of reference history resulting in wasting recency (temporal ist. LFU’s weakness in partial aspect may trigger keeping objects too long, cache space. While LRU is ocality), LFU depends on frequency (popularity). To reduce cache load, some studies based on proposed an aging mechanism that will evict long-lived objects, such as LFUDA [6], GD 6], GDS [6] and GDSF [1, 7, studies showed that size- hit rate while frequency- based meth based methods obtained 8], Some a better a better ods achieved byte hit rate. Some methods occupied more complicated techniques using statistical approach for data hooding. The typical one is the Lowest posed by Lorenzetti et al. model to calculate the re Relative Va [20]. The L ative value ue (LRV) policy, pro- RV uses a cost/benefit of each document in the cache, i.e., (CXP)/B, where C is the size of the requested document, expressing the overhead of ument from the original site, B is t transferring the doc- he bandwidth toward the server, and P is the probability that a document will com ful LWU is regarded as d of t to both recency and Anyway, the drawbac of cached documents. than LFU, LRU and other size-based be accessed again. In general, the LRV performs better k of As another we Least Unified-Value ( Usage A a model-driven simulation. The ynamic and sel mance seems better than ot techniques, but with plicated mechanism. However, it is particularly use- for small cache. For a larger cache, its performance drops. The Least Weighted icy was proposed in [23] as gorithm (LWU) pol- f-adaptive. Its perfor- her caching algorithms in terms he hit-rate. LWU can perform well with its adaptation frequency in di LWU is that it ignores the size fferent environments. 1-known method, the LUV) policy was proposed by Bahn
This subsection presents how we construct a set of fea- tures from the web log for ca culating replacement priority. Table 2 summarizes the features used for classification as a list. In our method, given a currently focused access request rj in the access log, we divid two regions preceding the req e a web log into five regions; uest (P; and P2), the region of the currently focused access request r;(C), and two regions succeeding the request (Ny and N2) as shown in Fig. 3. Twelve features (F' — Fg and F\o — F 3) are extracted from
This subsection presents how we construct a set of fea- tures from the web log for ca culating replacement priority. Table 2 summarizes the features used for classification as a list. In our method, given a currently focused access request rj in the access log, we divid two regions preceding the req e a web log into five regions; uest (P; and P2), the region of the currently focused access request r;(C), and two regions succeeding the request (Ny and N2) as shown in Fig. 3. Twelve features (F' — Fg and F\o — F 3) are extracted from
Table 2 List of features used for classification the first (P,), second (P2) and third regions (}), i.e., four features from each region. In addition, one feature related to the size (F9) is extracted from the fourth region(C). The target feature (F\4) represents the existence of the revisit of the request r; in the fifth region (V2). In total, there are fourteen features extracted for each access URL in a web log as a record, representing the request r;. In more details, the object repetition features (F1, F's, Fo) represent frequency of accessing the cache object which has the same URL with the currently focused request r; in the regions P2, P;, and
Table 2 List of features used for classification the first (P,), second (P2) and third regions (}), i.e., four features from each region. In addition, one feature related to the size (F9) is extracted from the fourth region(C). The target feature (F\4) represents the existence of the revisit of the request r; in the fifth region (V2). In total, there are fourteen features extracted for each access URL in a web log as a record, representing the request r;. In more details, the object repetition features (F1, F's, Fo) represent frequency of accessing the cache object which has the same URL with the currently focused request r; in the regions P2, P;, and
by t fold he SplitTrainTest function. Here, a standard k- cross validation can be applied for evaluation. The training dataset TR will be used to learn the model (Learnt- Mod (TS) accuracy (Accuracy) by the Tes is utilized to test the mod sub- GenerateN1Feature, GenerateN2 el) by the Training function and the test dataset el (LearntModel) for its ting function. The four functions, i.e., GenerateP 1 Feature, GenerateP2Feature, Feature in the Generate- DataSet, generate the features for the four regions (P}, P2, Nj, and N2). The size function returns the size of the object at the current URL.
by t fold he SplitTrainTest function. Here, a standard k- cross validation can be applied for evaluation. The training dataset TR will be used to learn the model (Learnt- Mod (TS) accuracy (Accuracy) by the Tes is utilized to test the mod sub- GenerateN1Feature, GenerateN2 el) by the Training function and the test dataset el (LearntModel) for its ting function. The four functions, i.e., GenerateP 1 Feature, GenerateP2Feature, Feature in the Generate- DataSet, generate the features for the four regions (P}, P2, Nj, and N2). The size function returns the size of the object at the current URL.
Here, p(Fi4 = +1|F;) is the probability that the record r (a web object) will be revisited again in the region N2, p*(Fi4 = +1|F;) is the unnormalized probability that the record r (a web object) will be revisited again, (i.e., positive class), and p*(F\4 = —1|F;) is the unnormalized probabil- ity that the record r is not revisited again in the region N2 (i.e., negative class), p(Fi4 = +1) (or p(Fi4 = —1) is the prior probability that the record is revisited (or not revis- ited). p(F; = rilFia = +1) (or p(F = rilFia = —1)) is the probability that the feature (F; = r;) when the class is positive (or negative). For continuous attributes (F;), Guas- sian distribution with smoothing can be applied as follows:
Here, p(Fi4 = +1|F;) is the probability that the record r (a web object) will be revisited again in the region N2, p*(Fi4 = +1|F;) is the unnormalized probability that the record r (a web object) will be revisited again, (i.e., positive class), and p*(F\4 = —1|F;) is the unnormalized probabil- ity that the record r is not revisited again in the region N2 (i.e., negative class), p(Fi4 = +1) (or p(Fi4 = —1) is the prior probability that the record is revisited (or not revis- ited). p(F; = rilFia = +1) (or p(F = rilFia = —1)) is the probability that the feature (F; = r;) when the class is positive (or negative). For continuous attributes (F;), Guas- sian distribution with smoothing can be applied as follows:
Table 3 Characteristics of three datasets: bo2 and sj during a wider time period of 10:15 to 15:48 with the re-reference ratio of 0.23 and 4.00 requests per second. trend. That is, the larger the N2 is, the larger the
Table 3 Characteristics of three datasets: bo2 and sj during a wider time period of 10:15 to 15:48 with the re-reference ratio of 0.23 and 4.00 requests per second. trend. That is, the larger the N2 is, the larger the
Fig. 7 Hit Rate (HR) comparison among LBR, GDSF, LFU and LRL (sj dataset) For the bo2 dataset with the cache size of 10 KB, our LBR outperforms LFU and LRU with the hit-rate increment of 1.22 % (39.98 % vs. 38.76 %), and 2.24 % (39.98 % vs. 37.74 %), respectively. They are 1.85 % (36.12 % vs. 34.27 %) and 4.00 % (36.12 % vs. 28.12 %), for the sj dataset. Moreover, for the bo2 dataset, when the cache size
Fig. 7 Hit Rate (HR) comparison among LBR, GDSF, LFU and LRL (sj dataset) For the bo2 dataset with the cache size of 10 KB, our LBR outperforms LFU and LRU with the hit-rate increment of 1.22 % (39.98 % vs. 38.76 %), and 2.24 % (39.98 % vs. 37.74 %), respectively. They are 1.85 % (36.12 % vs. 34.27 %) and 4.00 % (36.12 % vs. 28.12 %), for the sj dataset. Moreover, for the bo2 dataset, when the cache size
Fig. 6 Hit Rate (HR) comparison among LBR, GDSF, LFU and LRU (bo2 dataset)
Fig. 6 Hit Rate (HR) comparison among LBR, GDSF, LFU and LRU (bo2 dataset)
Fig. 8 Byte Hit Rate (BHR) comparison among LBR, GDSF, LFU and LRU (bo2 dataset)
Fig. 8 Byte Hit Rate (BHR) comparison among LBR, GDSF, LFU and LRU (bo2 dataset)
Fig. 9 Byte Hit Rate (BHR) comparison among LBR, GDSF, LFU and LRU (sj dataset)
Fig. 9 Byte Hit Rate (BHR) comparison among LBR, GDSF, LFU and LRU (sj dataset)
Related topics:
Computer ScienceDistributed ComputingOther Engineering
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved