Analysis

posted on 30 May 2008 16:45

HP's new 9100 Extreme Data System has a secret internal switching fabric.

The ExDS9100 is a stand-out winner in the scale out network-attached storage (NAS) stakes, with multiple petabyte scalability through one management interface. Performance block CPU blades and storage blocks are publicly defined yet its internal switch fabric is a mystery.

The server blades use x86 processors in a C-class chassis and there can be up to 16 blades with 128 CPU cores in total.The storage building block is a 7U enclosure with 82 1TB 3.5-inch nearline serial-attached SCSI (SAS) drives. Each storage block has two I/O paths for redundancy. Multiple ExDS9100s can be aggregated together to provide a multi-petabyte resource. The software is based on the HP PolyServe clustered file system.

Any server blade can talk to any storage block, with HP stating there is 'one pool of file storage with fully parallel symmetric access from all blades'

That means that there are could be point-to-point connections beteeen every server blade and every storage block. This is a recipe for a cabling nightmare and difficult additions of both server blades and storage blocks. The system has, HP promises, quick and easy upgrades of both performance and capacity. HP also tells us that you can 'add a storage block to the system by connecting two cables.' These points should rule out a P2P connectivity scheme.

The alternative is a switching fabric, similar in overall connectivity to a Fibre Channel SAN fabric, with additional server blades and storage blocks (pictured left) connecting to a fabric switch and software being used to add new resources into the system.

HP says the base configuration has four server blades and three storage blocks.

When asked about the switching fabric ineeded inside the system, HP StorageWorks VP for EMEA, Neal Clapper, replied: "... yes we need something to connect the servers and the disks. However, we are not able to divulge how we are doing this since we are too far from product release to give these types of details."

Product release is in the final quarter of this year, thee to six months away.

Logically, either HP has announced a base configuration of four server blades and three storage blocks, each with  two I/O paths, without  knowing what the internal switching fabric is going to be, or it wants to keep the switching fabric details secret, for a while at least. It has to be the latter as the former would be an announcement stopper.

Why might that be, why is there a need for HP to keep the details secret?

HP has Tandem's non-stop switching fabric available to it. It could use Fibre Channel and it might use an InfiniBand scheme. It might even be a new switching scheme altogether. The system is meant to be extremely affordable so the internal switching fabric cannot be expensive.

We can deduce something about the switch fabric structure. There can be up to ten storage blocks, each holding 82 1TB SAS drives and having two I/O cables. So ... there must be 32 ports for storage blocks.

There can be up to 16 server blades which could each hold dual quad-core Xeons or Opterons to give us the 128 core maximum. Let each CPU have its own port and we have 40 CPU ports in the fabric. Add in two management ports and we arrive at a 74-port fabric that has to serve 32 quad core CPUs and keep them busy talking to a maximum of 820 SAS I (3Gbit/s) disk drives.

Then there is the matter of external connectivity, presumably gigE. This is turning into a fairly chunky switching fabric requirement.

Server blades slide in and are physically connected automatically. Storage blocks are slid in and need two cables connected. To what? There could, literally, be a backplane in the ExDS8100 rack which hosts the switching fabric.

^{[Chris Mellor.]}