NANOG 08 feb 98 1400 kc presentatin

NLANR cache workload/performance graphs

NLANR return on investment analysis

9 feb 98

nlanr cache pi: Duane Wessels
data analysis: Alex Rousskov

return on investment (bandwidth)

cache or not to cache: small ISP

  1. ISP charged uniformly for outbound bandwidth.

  2. ISP always has unsatisfied demand for bandwidth (customers `waiting')

  3. income proportional to available bandwidth (new customers same income than old)

  4. Squid on mid-size unix box can support 50 requests/s, 3.9 Mbps, or 2 T1s
    (~~0.9 Mbps for various Squid overhead; implies saturating link)

  5. constant byte hit ratio sustainable

cache or not to cache: small ISP

  1. ISP earns monthly income of $X.

  2. saving bandwidth can allow ISP to admit new customers, generate more income

  3. question: will investment in proxy equipment and supprt pay off and when?

goal: find Y (in months) such that:

income_with_caching(Y) >= income_wout_caching(Y)

(note: income_with_caching(0) = - equipment_cost, so initially ISP loses money)

with I = total monthly income without caching (current bandwidth)

Y*(I + I*BHR) - equipment - Y*cache.admin== Y*I
solve for Y

cache or not to cache: small ISP

realistic parameters per T1

squid cache boxcost = $10K
equipment = boxcost / 2 (pipes);
cache.admin = $25,000 / 12 (months)

months_to_payoff(equip,admin,bhr,income) = equip/(income*bhr-admin) = Y


cache or not to cache: small ISP

-axis: number of months per T1 (1.5MBps)
ISP with monthly income (2 T1s, no cache) of $32K and sustained hit ratio of 25%
would double its investment in about 2*6=12 months
(2 connections, 6 = 25%: double curve (x=32))

return on investment (time)

NLANR supercache latency savings:

average hit saves US about 1 sec, European users ~2 sec
total (crude) NLANR supercache savings in response time:
( sec_saved_per_hit * #hits )

assume investment of 1 hour/day (365 days/yr) to monitor cache hierarchy

profit: time saved so far

ROI = (profit - investment) / investment

9 feb 98, alex/kc,