Fusion-io: a shot in the arm for server farms

Even so, there was room for improvement in this strategy. On the sort of older PC that’s commonly encountered teetering on the rim of a skip (the sort it gives me most pleasure to rescue), both IDE channels pass through the same, rather dim-witted controller chip, and trying to push two data streams through this would soon reveal its cheapness of spec and poverty of brainpower.

Fusion-io: a shot in the arm for server farms

That led me naturally to try fitting bargain-basement secondary IDE cards, which very quickly brought me to understand that:

1) lots of PCI IDE add-on cards have shockingly bad chipsets themselves; and

2) almost no PCs with expansion slots actually offer free and unrestricted access to the cards sitting in them. You always have to turn off something else; I usually disable the serial ports, and then kick over the pre-existing plug-and-play allocation.

So, it’s always been possible to make a PC, working as a server or as a solo machine, go quite a lot quicker by fitting more than one disk, provided you avoided some well-known pitfalls and chose your technology carefully.

March of technology

I’d like to be able to say at this point that the march of technology has made this body of hard-won wisdom redundant, so that if you find your single-user, single-disk SATA setup a bit too sluggish you can take advantage of the two, four or more extra SATA connectors on your motherboard, and these would reveal all sorts of opportunities to parallelise your disk I/O.

I can see that enforcing the use of a smaller boot partition is just about the only way to guarantee a decent performance window

Sadly, though, it isn’t that simple. As far as I can tell, the recent huge increases in disk capacities and data volumes have more than offset any increase in speed that SATA offers, so in many cases the performance of multi-disk SATA setups will feel slower than that of older machines.

I used to criticise those manufacturers that only format part of their disks to disguise the variety of hard disk models they were fitting; now I can see that enforcing the use of a smaller boot partition is just about the only way to guarantee a decent performance window, before your machine starts to slow down because it has to manage a quarter of a million files scattered all over a single 2TB hard disk.

Answers

There are some answers to this problem already out there. It’s always been possible to cram an expensive server box full of disk controller cards and then run disks outside the primary CPU case attached to these.

However, this leads to machine specifications for big corporate server rooms that many mid-range companies refuse to believe are worth the money: one box that costs £15,000 while containing almost no internal drives, another £10,000 spent just on RAID cards, each of which runs a metal shed-full of drives worth around £4,000 each.

That’s so far away from most people’s experience of a “bigger server in just one box” that credibility is lacking. Mid-range companies have a hard time believing that this sort of setup is anything but corporate pork-barrel madness, but the truth has always been that to service 20,000 users sharing a single server on traditional corporate Windows-based jobs, this is the way your architecture has to go.

Looked at from operating system level inside that hypothetical giant corporate operations centre, these mid-range guys become even more incredulous: all that cash spent and all you have to show for it at OS level is just a few extra drive letters! “Hell, we can do that!” they snort.

What they missed, and quite amazingly continue to miss even nowadays, is that these separate drive letters in such a massive heavy-metal server are merely the surface presentation of a completely separate I/O channel that has no dependency on the other activity going on in the OS.

Real speed

And this is where the real speed comes from – and surprisingly, I’d have to say that this is rather better understood in the Windows PC business sector than anywhere else.

Sure, Linux can be set up to take advantage of such parallel I/O architectures, but Linux people have a rather distracted, airy-fairy outlook when it comes to this kind of hardware cheating.

Disclaimer: Some pages on this site may include an affiliate link. This does not effect our editorial in any way.

Todays Highlights
How to See Google Search History
how to download photos from google photos