Sun and Greenplum's Overstated Appliance Throughput Numbers
I was digging into the technical details of Sun's "Thumper" data appliance (a dual-CPU dual core rack unit that can hold a massive 24 TB of data) and ran into the overblown claims problem that seems rampant in all vendors selling into the high end of the market.
In this case it's overstated system I/O throughput. Here's the clip straight from Greenplum's site on the Bizgres MPP/Sun deal:
If you take their claims at face value, scanning 1 TB of data in 60 seconds at low cost is an impressive achievement. The problem is that to get there requires a serious hardware investment and not just one 4-core unit . The facts are pretty simple:
- Scan 1 Terabyte of data in 60 seconds.
- Leverages the first and only data server that combines a 4-way server with 24TB of storage in a single integrated system.
The documented I/O rate is 2 GB/sec from disk to memory for one 4 core unit (third item under "at a glance")
1 TB of data at 2 GB/sec means one unit can process 1 TB in 500 seconds, or 8 minutes 20 seconds. That's a lot more than 1 minute.
We need a little over 8 units (16 GB/sec) to achieve a TB/minute scan rate.
If we use the $32,995 price (from Sun's online store) for a 12 TB unit rather than the 24 TB unit mentioned above, the cost will be $263,960. The 24 TB unit mentioned above would set us back $559,960.
That's a lot of juice to get a 1 TB scan rate. The problem with this marketing message is that they are trying to mix a performance claim at the high end with a capacity claim at the low end. I could get the 1 TB scan rate with a rack of white-box opteron PCs and Bizgres MPP and come in a lot lower than $263,960. All I would give up is about 100 TB of additional capacity.
This is why you should always read the specs before taking performance claims at face value.
Posted by Mark Monday, August 14, 2006 9:58:00 AM |