Diagonal scaling is making a site faster by removing machines. The term was created by John Allspaw, the operations manager of Flickr, who replaced 67 dual-cpu boxes with 18 dual quad-core machines and recovered almost 4x rack space and reduced costs by about 50%
Horizontal scaling (scaling out) means adding or removing client workstations with only a slight performance impact. Vertical scaling (scaling up) means migrating to a larger and faster server machine or multiservers.
See a video by CIO of CNET, on scaling out and don’t miss the 8 fallacies of distributed computing… and the 9th one!
The Law of Diminisging Returns, referred to also as variable factor proportions, states that
as equal quantities of one variable factor are increased, while other factor inputs remain constant -ceteris paribus- a point is reached beyond which the addition of one more unit of the variable factor will result in a diminishing rate of return and the marginal physical product will fall.
I think the reason I was associating these 2 is because I thought: In Allspaw’s model of diagonal scaling, this law is not only undermined, but is infact inverted.
Illustration of scaling up and out from the microsoft web applications libraries
Tags: CNET CIO, diagonal scaling, distributed computing, Flickr, horizontal scaling, John Allspaw, law of diminishing returns, scaling out, scaling up, variable factor proportions, vertical scaling
June 28, 2008 at 9:19 am
Diagonal scaling is an excellent way to mitigate efficiency issues in a data center; however, it does not undermine or invert the Law of Diminishing Returns. The key to understanding this theory is perhaps to restate it as the Law of Diminishing “Marginal” Returns (LoDMR). I’ll present two perspectives on this point:
(1) One application (in a data center environment) would be the increase of either vertical scaling or the increase of horizontal scaling – while holding all else equal. In either case, LoDMR definitely holds. Based on your post, I’m assuming you will agree with this assertion.
(2) Now let’s suppose we apply diagonal scaling to the LoDMR model as a ’single unit of input.’ (NOTE: I’m abstracting a little, of course, but this holds regardless of any arbitrary measure of Diagonal Scaling, so it doesn’t particularly matter).
Anyway, even with the incredible scale gained from Diagonal Scaling, at some point you will still experience “diminishing marginal returns.” If you exceed the optimal quantity of load balanced, multi-core machines that you need to support your environment, then the actual “marginal/variable” return you are deriving from each additional unit of “diagonal scaling” will in fact diminish – as in the classic labor application of the LoDMR, under-utilized resources are wasted resources.
I do, however, appreciate the point you are trying to make, but remember that there is always a ceiling to hit. In this case, it is not really a question of the LoDMR, but rather a question of efficiency. Diagonal scaling achieves far greater efficiency of resources than either horizontal or vertical scaling alone.