One or two 1gig nics on a 10g backbone sound reasonable with "only" 4 1T drives. 12*2T disks per node are getting more common and do not all have 10gig network cards, even on 600+ node clusters.
Sent from my iPhone
On Dec 23, 2011, at 11:15 AM, Mads Toftum <[EMAIL PROTECTED]> wrote:
> On Fri, Dec 23, 2011 at 12:23:59PM -0500, Koert Kuipers wrote:
>> For a hadoop cluster that starts medium size (50 nodes) but could grow to
>> hundred of nodes, what is the recommended network in the rack? 1gig or 10gig
>> We have machines with 8 cores, 4 X 1tb drive (could grow to 8 X 1b drive),
>> 48 Gb ram per node.
>> We expect "balanced" usage of the cluster (both storage and computations).
> If it were me, I'd start by looking at the amount of data you need to
> handle. Guestimate on both peak and average size. For simplification, start by
> guessing whether you'll be reading or writing more. If writes will be
> the driver, then you need to adjust for the number of copies. Or if you
> don't know how much data, then you could do a read and write test on one
> of the machines and figure out how much it can produce and go from that.
> I haven't looked at the pricing of 10gig gear lately, but I wouldn't be
> surprised if it made more sense for you to run simple nodes with a
> couple of 1gig cards and your infrastructure / backbone on 10gig.
> There's really too much ymmv based on your specific case, so that you'll
> only really know if you test on a handfull of machines.
> Mads Toftum