Oracle Open World 2008 Diaries: HP Oracle Database Machine

For those of you who didn’t see the Larry Ellison’s keynote here it is courtesy to Sheeri.

We cut out the HP part but I don’t think anyone will complain. It’s not the best angle but we didn’t get there early in advance to secure the right location for the camera.

Now few bits about the details of the HP Oracle Database Machine (HPODM) as I have gathered some details here at OOW. Oracle worked hard in two areas — perfection of hardware and software integration. Every layer is thought of and engineered for IO performance. Oracle worked closely with HP engineers and lots has been done in the HP firmware to make sure there are no bottlenecks. This will also help OLTP applications a lot. Each storage cell contains 12 disks and they can deliver 200 random IO’s per second from the outer “hot” area (55% of the disk surface) assuming “cold” areas are not used. That gives us 2,400 random IO’s per second (IOPS) or 33,600 IOPS for the whole Database Machine with 14 storage cells. I think there might be thoughts to have storage cells configured with flash drives as well — it’s standard 3.5″ disks. Throughput figured for sequential IO are on Kevin Closson’s blog.

About the storage cell itself — it basically a Linux box with internal hard disks — HP ProLiant technology. Is there an Oracle instance running on the storage server just like ASM? No. Oracle software does include few libraries from Oracle database installation but it’s just an RPM package. I’m so glad Oracle went this way — I wish Oracle software will be deployed this way.

The smart part includes filtering predicates processing on storage side, filtering of columns sent to the database server, and assistance in joins. The joins are not done entirely on the storage server but it does the part of bloom filtering algorithm and eliminates significant part of the rowset that would otherwise be sent to the server in full. There is no sorting on the server side or at least not yet. I didn’t look into the details of query results caching on the storage server so no comments on that.

What’s important to note is that table compression is fully supported — Exadata Storage Server (which is what a storage cell is) is aware of compressed block structure (remember I mentioned few database libraries are used?).

What about small customers? Exadata Storage Server is available separately. I think that soon hosting services will be available based on HPODM and Oracle VM. It’s just that Oracle did all the difficult bits of integration — no problems with compatibility of all components and setup efforts are minimal. We just need to make sure that licensing can be done efficiently. What about very very large customers and databases — HPODM cabinets can be stacked together with Infiniband.

OK. I’ve got to run now. Need to say goodbye to many good friends here in San Francisco.

PS: Oh… I forgot to say that HPODM runs RAC, of course. What else? There are 8 “normal” server units in a single HPODM but I think that’s flexible. I wonder if you can license Standard Edition for them? There are 2 quad core CPU’s.