Build a performance model that makes us of the CBMG data of last week.
Problem: number of auctions and bids surges from 8pm - 11pm.
Approach: analyze performance for each type of request during the
peak period, as number of sessions increases. Note: current SLA specifies
max R of 4 sec for creating new auctions and viewing bids.
Build performance model using multi-class queuing network: 6 classes
(h, s, v, g, c, b) and 6 queues (WS_cpu, WS_disk, AS_cpu, AS_disk, DB_cpu, DB_disk).
To solve, need per-class arrival rates, and per-class, per-device service demands.
To calculate per-class arrival rates (lambda_x, x in {h,s,v,g,c,b}),
need overall arrival rate (gamma), f_A (fraction of users in cluster A),
f_B (fraction of users in cluster B). Then:
lambda_x = gamma * (f_A * V_x^A + f_B * V_x^B)
The calculated arrival rates can be found in the same spreadsheet, on
the tab marked "Workload-Intensity". Assumputons are gamma = 10.96, f_A = 0.25, f_B = 0.75.
Still -- need the service demands. How to get them? Create scripts
that repeatedly submit a single type of request. Then:
submit N (large number) of requests one type to the site during time T
measure U_i at each device during that period
apply service demand law ( D = U/(N/T)).
Calculate R_r for each class r, using
R_r = Sum(i=1 to K) D_i,r / (1 - U_i)
Note: U_i will vary with the arrival rate of requests.
U_i = Sum(r=1 to R) U_i,r = Sum(r=1 to R) lambda_i,r * D_i,r