Better 1x10 than 10x1 (part II)

This is a mirrored from: ibm.biz/demystperf.

Here are two additional graphs illustrating the fact “the bigger, the better”.

Here we are in the same scenario considered in the previous “Better 1x10 than 10x1” blog post, introducing a new XXXL server, 10 times bigger than the XL one.

The first chart illustrates the usage of the servers when loaded with a customer population that fill the server to an average value of 50% of its capacity, that is, 10 users for the 10x sized (M) server, 100 users for the 100x sized (XL) server and 1000 users for the 1000x (XXXL) sized server.

The green zone is where the usage resides 50% of the time. The yellow zone is visited 40% of the time. And the server is in the remaining white zone only the 10% of the time. It is visually clear that the more users, the more close to the average the usage wanders (the relative variability reduction).


  

The second chart shows the result of a much better sizing: reduce the almost deserted white zone to begin at 90% usage. How many users do the servers support with this sizing target?  In another words: how many users push the 10% of the time zone to 90%-100% usage? You agree this is clearly a better sizing,  don’t you? Here are the results:

Answers:

• You can load the 10x sized server with 13 users, driving up the average usage to 65%.
• You can load de 100x sized server with 164 users, driving up the average usage to 82%.
• You can load the 1000 sized server with 1746 users, driving up the average usage to 87%.

A technical remark: I’ve obtained the above values with the binomial distribution. When the number of users is “big enough”, the binomial approaches to the well known normal distribution. Stop for the moment; perhaps I’ll come back on this in a later post.