Skip to main content

Your browser may not be compatible with all the features on this site. Consider upgrading to a modern browser for an improved experience.

View Post [edit]

Poster: Curator at the Security Digest Archives Date: Feb 17, 2005 8:02am
Forum: petabox Subject: Re: about 400TB of this design shipped to the Archive

You need some custom hardware, it seems to me: i.e. a DIMM-PC or gumstix style form factor. Essentially, it needs to be the same size as the 3.5' form factor so that it can fit snuggly onto the back of your IDE drive. The hardware needs only a 200+mhz processor, 64mb ram, ethernet and IDE controller; it's not a tough call to do this on that size of PCB. The hardware could run uLinux or NetBSD to export CIFS, NFS, ATA over Ethernet or whatever you require - effectively a DIMM size SAN. This type of hardware would be low power, give off little heat, be cheap and reliable. Not only that, I'm sure if you over-produced and sold some of them off, you'd be able to recover some cost. The VIA boards are great, but the real estate, MTBF and power drain would really have to add up at petabox scale.

Reply [edit]

Poster: Brak Date: Feb 17, 2005 1:23pm
Forum: petabox Subject: Re: about 400TB of this design shipped to the Archive


You make a good point and I've often thought of something similar but it's not cost effective yet.

A good way to look at the equation is the delta per block device. Being vague about block device since sizes and type are many. The services you need to provide are access and power at a minimum and perhaps processing, more on that later.

Access ~ Megabits/Block Device, switching (level of oversubscrition on uplinks, if any), cabling, etc.

Power ~ Cabling, Power supplies, etc..

Cooling ~ putting thousands of these next to each other is more challenging that putting one on your desk.

A desktop PC version could be as cheap as a sempron board with CPU and a desktop powersupply and a linksys switch.

Stick it into a spreadsheet, play around..

The upshot in all this is processing. Once the scale is hundreds of nodes, and thus thousands of disks, the incremental cost, if any, of real CPU's, when compared to other models can be quite compelling.

Once you have all the variables plugged in, you'll notice the costs of things like having 4 100Mb ports vs. 1 1000Mb port. You'll also see how choosing between 4 Arm CPU's, 1 VIA or a Xeon affect your monthly power bill (don't forget about the AC.)

Even after all of that, some kind of erector set needs to suspend all this stuff (maybe in cases, maybe not) in some kind of room... keeping everything cool.