In this paper I will show you how to perform a Capacity Planning (CP) of CPU and I / O in a simple manner using simple math. To make an effective CP is important to determine the sub-modeling. The subsystems are typical CPU, I / O or disks, memory and network. There are commercial products that allow a global PC without having to model for the subsystem, but I prefer to analyze each subsystem separately because I think that the results are more precise and clear. The CPU and I / O are clearly the most important are the subsystems in which we delve into this note. Surely
once your manager has asked them if the number of processors or the disk speed is adequate to ensure that the system remains stable. To this question one can answer just be confident and everything will work well or one can be more professional and do a PC and know exactly where it will be the breaking point, ie not only answer if the HW supports the current load but also, taking the annual growth forecast transacccional, to know where we reached our HW and predict in advance the purchase or upgrade of HW.
now go directly to see how modeling a CPU subsystem and a subsystem of i / o. To perform more sophisticated CP must have some knowledge of queuing theory (Erlang C function, Kendall notation, etc). In this note simply achieved with a little discretion. Figure A) is modeled CPU subsystem. This consists of a single queue and N servers (CPU's). This means that each request can be served by any server arbitrarily or gluing cpu and go when all the CPU is busy (runqueue in terms of OS). Figure B) is modeled subsystem I / O where we have a queue per server or device. In this case, each petition must be served by a particular device.
λ= tasa de arribo (ej: trx/ms ó Uc/ms)
The two models mentioned are represented many times in daily life, when we go to the bank, when we are waiting at the toll booth, as we wait to board a flight, when we place our order in a restaurant de comida rapida, etc.
Ahora vamos a definir las variables que necesitamos usar para realizar el CP:
<= 100000
Uc= User Calls
trx= transacción
St= Tiempo de Servicio (ej: s/trx ó s/Uc)
Qt= Tiempo de Espera o Tiempo de espera (ej: s/trx)
Rt= Tiempo de Respuesta (ej: s/trx)U = St * λ / M
Q= Encolamiento (ej: Nro de trx's)
U= porcentaje de carga (porcentaje de utlización del recurso)
M= nro de servers (cantidad de CPU's o cantidad de IO devices)
y las formulas:
Rt = St + Qt
Q = λ *
Qt Rt (cpu) = St / (1-U ^ M)
Rt (io) = St / (1-U)
The goal is to graph the response time vs arrival rate. The graph should be exponential with a curve or turning point that shows that arrival rate (X axis) occurs the break.
I will explain more in detail what each thing means:
1) λ (Arrival Rate): This metric can be defined with Oracle such as: Uc / s Trs / s, or using functional metrics, eg purchase order charge per minute or per hour.
2) M (Quantity of CPU'sy I / O Devices): The number of CPU's or I / O Devices. To obtain this value we can use OS utilities and / or consult Oracle views.
3) U (Percent Charge): The percentage of CPU load or I / O. This can be achieved using OS utilities such as sar, iostat.
4) St (service time): The time in which the request is being handled by the CPU or the device i / o.
5) Qt (Waiting Time Tail): The time that a request has to wait in line because the servers are busy.
6) Q (Long Tail): The number of requests queued (waiting to be served)
7) Rt (Time response): The response time is the sum of St (service time) and Qt (timeout or bonding). A request or is waiting or being served.
At this point, where I have defined all, I'll show you an example: Example
(Capacity Planning CPU) <= 100000
Suppose we are asked to estimate the impact that increases will occur if the user of an application X by 20%. What we want to determine is whether you need to add more processors or what is good enough to support the additional load.
first thing we do is collect information on a representative period. The more data we collect more accurate will be our CP, then we must characterize the load, this means defining whether we will use average values, maximums, which we will use as arrival rate, etc. In the example I'm going to have peak values \u200b\u200bin the range of maximum load and a rate up using Trx / s.
λ = 20 trx / s (20 transactions per second, is available from Statspack or AWR)
U = 0.40 (40% cpu utilization, it can be collected with sar-u)
M = 8 ( there are 8 processor then, is obtained from the parameter of the base: cpu_count)
Since I have the 3 essential parameters, only now I have to apply the formulas and graph:
St * U = λ / M, depejando for St, St
= U * M / λ = 0.40 * 8 / 20 = 0.16 s / trx
St
Now we can apply the formula for cpu:
Rt (cpu) = St / (1 -U ^ M) = 0.16 / (1-0.40 ^ 8) = 0.16 s / trx
As seen the response time equals the time of service. This is because the system is loose in CPU for the current load and there is no queuing, ie wait to be served by the cpu's. Now what I do is a table in excel to project and plot the growth curve in order to see where the break occurs.
see that the turning point is around 30-35% (>), consider also that we are starting from a peak workload, and therefore we think that most of the time will be well below. According to CP can be sure that the damper system while the additional burden of 20%, beginning to deteriorate rapidly from 25%. Example 2 (Capacity Planning for I / O)
5Mb/ms λ = (5 Mb transfer per millisecond can be obtained with sar-do iostat)
U = 0.60 (60% usage of i / o)
M = 50 (there are 50 devices)
St = U * M / λ = 50 * 0.60 / 5 = 6 ms / Mb
Replacing in the formula i / o: Rt (io) = St / (1 -U) = 6 / (1-0.60) = 15ms/Mb
As shown in the graph, the system could support up to 50% (7.5Mb/ms) of growth, once reaches 60% of destabilizes.
In summary we can conclude that our hardware is well at both non-cpu i / o as long-term will not scale. This paper
my idea was to show how to make a cpu capacity planning yi / or easily. With this method we have a good estimate how to escalate our HW. According to each case, there are other approaches such as modeling method based on ratios, queuing theory and linear regression method.
most important for accurate estimates is to obtain a sample that is representative joined a good characterization of the workload to be assessed. The more information we have collected better prognosis.
Using this methodology can perform simulations and future projects, for example, what if we add or we get cpu's, how it impacts the aggregate device i / or faster, with higher throughput and lower latency, few users could add to operate with application without compromising system stability, etc.
To those of interest issues in Oracle Capacity Planning I recommend:
"Forecasting Oracle Performance"
. This book by Craig Shallahamer, a real guru on the subject, was excellent and I used it as reference to write the note.
select dbms_stats.get_prefs ('INCREMENTAL', tabname => 'T') from dual; 
dbms_stats.gather_table_stats (ownname = > user, tabname => 'T');