Evaluating Content Management Techniques for Web Proxy Caches Martin Arlitt, Ludmila Cherkasova,...

Evaluating Content Management Techniques for Web Proxy Caches

Martin Arlitt, Ludmila Cherkasova, John Dilley, Rich Friedrich and Tai Jin

Hewlett-Packard Laboratories

4th International WWW Caching Workshop

元智大學資訊工程所 系統實驗室陳桂慧

1999.10.06

Outline

• Key workload characteristic• Experimental design • Simulation results• Conclusion

Key Workload Characteristics

• Cacheable objects• Object set sizes• Object sizes• Recency of reference• Frequency of reference• Turnover

Experimental Design

• Cache size– 256 MB, 1 GB, 4 GB, 16 GB, 64 GB, 256 GB and

1TB…...

• Cache replacement policy– LRU, SIZE, GD-Size, LFU, GDSF, LFU-DA

– LAT, HYB

• Performance metrics– Hit rate

– Byte hit rate

Replacement Algorithm (1)

• Least-Recently-Used (LRU)– replaces the object requested least recently.

• SIZE– replaces the largest object.

• LFU– replaces the least frequently used object.

• GreedyDual-Size (GD-Size)– replaces the object with the lowest utility.

– Ki = Ci / Si + L

Replacement Algorithm (2)

• GreedyDual-Size with Frequency (GDSF)– Ki = Fi * Ci / Si + L

• Least Frequently Used with Dynamic Aging(LFU-DA)– Ki = Ci * Fi + L

Hybrid Algorithm (HYB)

• Motivated by Bolot and Joschka’s algorithmW1rtti + W2 si + (W3 + W4 si)/ti

– ti : the time since the document was last referenced

– rtti : the time it took to retrieve the document

• (clatser(i) + WB/cbwser(i))(nrefi** WN)/ si

– nrefi : the number of references to document i since it last entered the cache

– si : the size in bytes of document i– WB and WN : constants that set the relative importance of the variables

cbwser(i) and nrefj

Latency Estimation Algorithm (LAT)

• clatj = (1-ALPHA) clatj + ALPHA sclat

• cbwj = (1-ALPHA) cbwj + ALPHA scbw.– Clatj : estimated latency (time) to open a connection to the

server– cbwj : estimated bandwidth of the connection– sclat and scbw : the connection establishment latency and

bandwidth for that document are measured

• di = clatser(i) + si/cbwser(i)– ser(i) : the server on which document i resides

– si : the document's size – di : LAT selects for replacement the document i with the smallest

download time estimate

• Comparison of existing replacement policies

GD-Size(1)LFU-AgingSIZELFUGD-Size(P)LRU

LFU-AgingGD-Size(P)LRULFUGD-Size(1) SIZE

• Comparison of proposed policies to existing replacement policies

GDSF-HitsGD-Size(1)LFU-AgingLFU-DAGD-Size(P)LRU

LFU-AgingLFU-DAGD-Size(P)LRU GDSF-HitsGD-Size(1)

Virtual Caches

• An approach that can focus on both of high hit rates and high byte rate.– each virtual cache (VC) is then managed with its own rep

lacement policy. • initially all objects are added to VC 0,

• replacements from VC i are moved to VC i+1,

• replacements from VC n-1 are evicted from the cache.

• all objects that are reaccessed while in the cache (i.e., cache hits) are reinserted in VC 0 .

– this allows in-demand objects to stay in the cache for a longer period of time.

GDSF-HitsVC-HB-75/25VC-HB-50/50VC-HB-25/75LFU-DALRU

LFU-DAVC-HB-25/75VC-HB-50/50VC-HB-75/25LRUGDSF-Hits

• Analysis of Virtual Cache Performance – VC0 using GDSF-Hits, VC1 using LFU-DA.

• Analysis of Virtual Cache Management – VC0 using LFU-DA, VC1 using GDSF-Hits.

GDSF-HitsVC-HB-25/75VC-HB-50/50VC-HB-75/25 LFU-DALRU

LFU-DAVC-HB-25/75VC-HB-50/50VC-HB-75/25GDSF-HitsLRU

• Analysis of Virtual Cache Management – effects of VC order on performance

VC-BH-25/75VC-HB-75/25VC-BH-50/50VC-HB-50/50 VC-BH-75/25VC-HB-25/75

VC-HB-25/75 VC-BH-75/25 VC-HB-50/50 VC-BH-50/50VC-HB-75/25VC-BH-25/75

Conclusion

• Size-based policies achieve higher hit rates than other policies.

• Frequency-based policies are more effective at improving the byte hit rate of a proxy cache.

• Virtual caches as an approach provide optimal cache performance for multiple metrics simultaneously.