Musing: Hyperscale computing and us

Fox, Kevin M kevin.fox at pnnl.gov
Wed Apr 10 18:53:15 UTC 2013


I believe HP has been contributing to https://wiki.openstack.org/wiki/Baremetal already. Basically using the same openstack api for bare metal instead of vm's. This allows all the same software to manage the instances whether physical or virtual. This hardware would fit that model very well.

Kevin
________________________________________
From: aeolus-devel-bounces at lists.fedorahosted.org [aeolus-devel-bounces at lists.fedorahosted.org] On Behalf Of Matt Wagner [matt.wagner at redhat.com]
Sent: Wednesday, April 10, 2013 11:50 AM
To: aeolus-devel at lists.fedorahosted.org
Subject: Musing: Hyperscale computing and us

There's been some buzz about HP's new "Moonshot" server:
http://h17007.www1.hp.com/us/en/enterprise/servers/products/moonshot/index.aspx

The main page is kind of marketing-speaky, but if you poke around it
becomes clear what it is: a 4.3U[1] enclosure that fits "up to 45
hot-pluggable, efficient, extreme low-energy servers." They're dual-core
2 GHz Atoms, with a single memory slot up to 8GB, and a single 2.5" disk.

I had lunch yesterday with some people who happened to be pretty
familiar with this, and what's dubbed the "hyperscale" paradigm. The
idea is that, instead of some dense servers and running VMs on them, you
go for a massive number of very low-power servers. These Atoms seem to
use something in the ballpark of 10W, but there's also a lot of interest
in ARM chips where you can get much lower usage. And apparently, 45
nodes in 4.3U is nothing. Look at HP's Redstone[2] for example, which
fit 288 ARM chips into a 4U chassis.

So on the surface, this is wholly irrelevant to cloud computing. In
fact, if anything, it seems like it undermines the whole premise of
virtualization.

But it occurs to me that, if it's possible to have hundreds, even
thousands, of servers in a single rack, the way people manage them is
probably going to change. I think you're almost going to want to treat
them like cloud instances -- boot them with some premade image, do your
work on them, and then spin them down, treating them as stateless. Maybe
it's a one-off compute job (using Hadoop or something), or maybe you're
hosting some sites and want to just power up additional servers as load
grows. I think you're going to want to stop treating them as always-on
servers with a fixed role, and view them exactly the same was as someone
views a cloud instance, but backed by hardware.

I think it's too soon to be able to do a lot, but I think this is
something worth thinking about. Can we effectively bring what we do in
the cloud to "hyperscale" physical nodes?

-- Matt

[1] Yes, 4.3U. And yes, the idea of fractional rack units really bugs me.
[2] http://www.theregister.co.uk/2011/11/01/hp_redstone_calxeda_servers/


More information about the aeolus-devel mailing list