Re: Larry Ellison comments on Microsoft's benchmark

In article <P3Ib5.43053$i5.567197_at_news1.frmt1.sfba.home.com>,   "Ivana Humpalot" <ivana_humpalot_at_nospam.com> wrote:

> I know that with DB2 on MSCS, you only have two
> options for HA: Either have a mutual takeover situation, or
> have failover. There are defects with both. In a mutual failover
> situation the entire load on the failed machine has to be taken
> over by just one machine. If that machine is already heavily
> loaded then DOUBLING the load on that machine will likely
> overwhelm that machine, potentially killing that machine too.
> In a failover situation you double the number of machines which
> increases costs, and you get poor utilization of resources
> because half of the machines are idle.
>

I agree in principle with what you said. Doubling number of machines clearly increases costs. So let's do simple arithmetic for the just published TPC-C benchmark:

Simply doubling the number of servers (from 32 to 64) and having them sitting idle, would raise the system price by about $2.5 million. If performance were to remain unchanged, then only price/performance would be affected, rising from $32 to $38. (Note that TPC rules prohibit making performance projections, this is just an "e-paper" exercise). This price/performance is still much better than the best reported by Oracle.

But even the above is too pessimistic. AIX supports 32-node clusters and Windows 2000 Datacenter Server will support 4-node clusters, so doubling of servers is not necessary.

> But even with DB2 on 32-way clusters, if you only have 1 idle
> machine, if 2 machines fail then the load on at least machine will
> double, which could lead to a total system failure. In Oracle
> Parallel Server, 16 machines will have to die before the load on
> one machine doubles.

"If 2 machines fail"? Nothing different from a question of "what if 2 disks in a RAID-5 array fail" -- you get system outage until one of the failed units is replaced. Inconvenient, but hardly catastrophic.

By the way, it's pretty hard (easy?) for 16 machines to die using Oracle Parallel Server, considering that there isn't much evidence that OPS can support more than 8 effectively :-)

> In a cluster, as you add machines reliabilty is supposed to go
> UP, not down. In Oracle Parallel Server, as you add machines
> reliability goes UP. In DB2 MSCS clusters, as you add machines
> reliability goes DOWN. Are you saying this is not a concern
> when adding machines?
>

I think there are two ways to look at clustering.

1. As you add additional servers, performance is supposed to go up. This has been demonstrated by DB2 with 32 nodes in this benchmark, and with customers with 100+ nodes. I haven't seen any data on Oracle with more than 8 nodes.
2. As you increase the number of servers in a clustered system, the system is expected to maintain its high availability characteristics without significant increase in expenses. I believe that all OS-based clustering solutions achieve this goal. While it is true that a multiple-node failure can impact cluster availability, it's not a concern to most customers. I suppose NASA can afford to have triple redundancy mechanisms, the rest of us are quite happy with double redundancy. If double-redundancy is good enough for disk subsystems (which are more critical than servers because they store non-volatile data), then it's good enough for clusters as well.

Gene

Sent via Deja.com http://www.deja.com/
Before you buy. Received on Fri Jul 14 2000 - 00:00:00 CDT