Cost-benefit analysis

Web Cache in UNINETT

Analysing the web cache system of UNINETT for efficiency, based on experience and statistics from installed servers.

Introduction to the subject is available in the document "Web caching: Costs and Benefits".

Cost model

The costs related to running av web cache system are

The benefits are

The costs for an Internet connection are

Both these costs are depending on the bandwidth. The cost of access lines are determined by bandwidth and distance, and a monopoly of access lines is held by Telenor until 1.1.98.

System

The current configuration (september 1997) of the web cache system in UNINETT consists of

Bytes saved

First level cache

The calculations are taken from www-cache.uit.no at Tromsų University. 

Month      To         From         Saved    Hitrate 
199701    16.72 GB    24.00 GB     7.28 GB  30.35 %
199702    16.51 GB    24.05 GB     7.54 GB  31.36 %
199703    15.49 GB    22.80 GB     7.31 GB  32.05 %
199704    19.45 GB    27.99 GB     8.55 GB  30.53 %
199705    16.60 GB    23.82 GB     7.22 GB  30.32 %
199706    12.95 GB    17.87 GB     4.92 GB  27.53 %
A smaller institution may have a web traffic of around 500 MBytes flowing in to the users per working day (measured a central point in network). Given 20 working days in a month, this is the equivalent of 10 GBytes per month. A typical byte hitrate (30%) gives some 3 GBytes saved traffic per month.

Top level cache

Total downloaded from www-cache.uninett.no (top level cache) to all children and neighbors 
Month      To         From         Saved    Hitrate 
199701    46.81 GB    55.31 GB     8.50 GB  15.37 %
199702    36.73 GB    54.94 GB    18.21 GB  33.15 %
199703    30.10 GB    59.68 GB    29.58 GB  49.56 %
199704    35.83 GB    69.09 GB    33.25 GB  48.13 %
199705    36.92 GB    71.40 GB    34.47 GB  48.29 %
199706    30.54 GB    46.77 GB    16.23 GB  34.70 %

Overhead due to ICP

The Internet Cache Protocol (ICP) is used for inter-cache communication, to determine which of the cooperating caches holds the requested document (and is also a load balancing feature).

The overhead per ICP-packet is around 64 B/packet. If June 1997 is taken as an example 17.8 million connections give 2.1 GBytes in added traffic due to ICP traffic (2 packets per connection with 64 bytes overhead in each.

Traffic generated by ICP is normally generated locally where traffic costs (in time and money) is less than it is for international traffic, and is in the order of 5% overhead per parent/sibling.

Total savings

The total savings are estimated for a typical month:
System savings per server total savings
top level 30 GB 30 GB
large first level 7 GB 21 GB
small first level 3 GB 36 GB
Total 87 GB

System savings raw bit rate
(Bytes divided by time)		 access line needed
(calculated over all hours)		 access line needed
(calculated over working hours)
top level 30 GB 50.1 kbps 125.3 kbps 1 Mbps
large first level 7 GB 21.6 kbps 54 kbps 243 kbps
small first level 3 GB 9.3 kbps 23.3 kbps 104 kbps
Total system 87 GB 145 kbps 363 kbps 2.9 Mbps

Raw bitrate is calculated as: (Savings * 8) / (3600 * 24 * 30) = bps

Access line needed (calculated over all hours) is calculated as 2.5 * (Savings * 8) / (3600 * 24 * 30)

Access line needed (calculated over working hours) is calculated as 2.5 * (Savings * 8) / (3600 * 8 * 20)

Latency savings

Time savings have not been converted to money saved, as it is difficult to detemine what the alternative use of time for each user is. Some users may download more web documents, some may use the time for other work related activites.

For the end user time savings are the most important benefit, as this is easily perceptible. The end users usually notices the addition of a proxy cache only by the reduction in download time.

A misconfigured web cache, or a web cache under heavy load may add latency.

latency

As the graph shows, hits are considerably faster than misses, saving on average 2 seconds in weekends and 4.5 seconds on working days per web object. Each web page is made up of several documents, normally one HTML document and 10-40 gif pictures. The total savings for a web page is 10-40 times the savings on each document. These estimates are done from the top level UNINETT server, where bandwidth to the US is plentiful. First level servers behind slow connections may have better results.

Savings calculated for hits only, which gives a lower bond on time savings. The same amount as hits will be for GET-If-Modified-Since, as shown on the figure above.
week end work day
seconds saved per object 2 4.5
hits/day 10 000 25 000
total time saved per day 5.5 hours 31.25 hours
Total savings per month 680 hours, 4.25 FTE

Cost of access line

SURFnet has done the calculations on basis of Dutch prices, UNINETT customers experience different national price for bandwidth.

The most realistic would be to use the price of 2Mbps as this is the nominal access capacity of UNINETT (256 kbps is the real thing, for most colleges). This gives a very much higher price. Doing the aritmethic with 34 Mbps lines gives a more favorable comparison, with only 4-10 times the NL price. Details on pricing of access lines in Norway, fall 1997.

Cost of access line used in calculations
Bandwidth 256 kbps 2 Mbps 34 Mbps
Median cost
for the equivalent of 1 Mbps		 1920 ECU 376 ECU 106 ECU

Cost to ISP

Payment to ISP from connecting institution covers national and international bandwidth and the administration of this infrastructure. The substantial part of this is for the international infrastructure that the ISP operates, due to the expensive international lines. In Europe this cost is higher as one needs to connet to the US by crossing the Atlantic.

The average price per month for 2 Mbps paid locally in Norway is 6000ECU, which is 3000ECU per month for the equivalent of 1Mbps.

Unless the actual cost of the international infrastructure is visible to the customer, the savings (in money) will not be visible.

Web cache server cost

The cost for web cache servers include both initial cost for a computer and yearly administrative cost in manpower (FTE).

Estimates of hardware costs. Hardware cost are depreciated over 3 years. Small sites may use a PC with Linux/FreeBSD , larges sites will need either large PCs or big UNIX systems
System Hardware, initial cost Manpower Total cost server/year
top level 18000 ECU 150000 NOK 0.50 FTE 28500 ECU
large first level 10000 ECU 80000 NOK 0.10 FTE 7800 ECU
small first level 2500 ECU 20000 NOK 0.05 FTE 3000 ECU
Total for system 87900 ECU

Conclusion

The use of web caches is not widespread in UNINETT today. Around 20 institutions out of 450 participate in the web cache system, even if some of these are among the larger institutions.

The savings for the local institution is depending on the actual cost for access lines from Telenor, and the type of server installed. The tabel below gives an overview of the savings in todays system.

System Cost server/month Saved traffic Equivalent leased line Estimated access savings Estimated ISP savings Estimated savings per month
Small first level cache 250 ECU 3 GB 100 kbps 192 ECU (256kbps eq) 300 ECU 242 ECU
Large first level cache 650 ECU 7 GB 230 kbps 87 ECU (2Mbps eq) 690 ECU 127 ECU
Top level cache 2375 ECU 30 GB 1 Mbps 106 ECU (34Mbps eq) 3000 ECU 731 ECU
Total system savings per month
(1 top level, 3 large first level, 12 small first level)		 4016 ECU (33kNOK)

Potential savings

More use of web cache and possibly blocking port 80 (for students?) will give higher savings. The table illustrates what may happen when web caching is widespread in the user community.
System Cost server/month Saved traffic Equivalent leased line Estimated access savings Estimated ISP savings Estimated savings per month
Small first level cache 250 ECU 4.5 GB 150 kbps 288 ECU (256kbps eq) 450 ECU 488 ECU
Large first level cache 650 ECU 35 GB 1150 kbps 435 ECU (2Mbps eq) 3450 ECU 3235 ECU
Top level cache 2375 ECU 90 GB 3 Mbps 318 ECU (34Mbps eq) 9000 ECU 6943 ECU
Total system savings per month, 4 large and 50 small first level 44283 ECU (363 kNOK)


Ingrid.Melve@uninett.no 2002-10-29