Re: Getting the best out of new server hardware: a disk setup for Oracle database

Mladen Gogala kirjoitti:
> Anyway, with your architecture, your bottleneck will be
> motherboard, not the disk array. PC servers are notoriously lousy at
> performing high volume I/O.

We have high hopes that the characteristics of our motherboard and architecture lands somewhere between a traditional PC server and an enterprise-class UNIX server.

> With proper quantity of NVRAM on your disk
> array, you will start seeing bottleneck of the architecture around 1200
> I/O requests per second.

We have managed to get 3987 IOPS out of our system (long Orion run, 8 kb random IOs, 100% reads, card caches set to 50/50 and 128 kb stripe size).

> To simplify things, on a PC-server, all I/O devices,
> CPU cache coherency, memory and video-card are on the same bus.

This is a little bit different on AMD architecture. Here we have HyperTransport bus to transfer the data between the processors and in/out from the IO system. This should be more efficient setup than the one with Intel Xeons, and was one of the reasons that we chose AMD over Xeon. And both processors have their own memory controller running at the same speed than the CPU cores.

Of course this is still far from "real" UNIX servers.

> Despite the marketing hooray about Linux servers, one always gets what
> one pays for.

I agree. But we are looking forward to see good price/performance ratio, not the ultimate performance readings regardless of the price tag. We are using the taxpayer's money.

> If you wish to entrust
> your data to ASM, don't say you weren't warned.

What I would appreciate are comments from those that have already used ASM on their production system. It may be a good policy to adopt new features slowly but what I want to know is how realiable ASM really is on a real-life production system.

--
Heikki