Every now and then we see people declare themselves to be the king of the hill in performance or in some other metric of relevance. Then come the theoretical max numbers or their measurements.
As a reminder, our single JR4 units are sustaining 2.3+ GB/s to and from disk for TB sized files, and have QDR IB type connections (as well as 10GbE and GbE) available. Lots of bandwidth per box. Far more, for example, than a recently self-proclaimed “worlds fastest” unit.
We did see another competitor show how it took 3 shelves of systems to hit 70% of the performance of a single one of our systems. It took them far more cost, and more complexity to achieve what we easily handle.
And then there was a video which reminded me of a whitepaper I saw someone else write up some years ago, on how they took 48 drives, in a jbod configuration, and enough RAID cards to drive them all, and then ran IOmeter on them. Work with me on this: assume 110 MB/s per drive, and in a non-RAID config, you can achieve ~5.2GB/s. In a configuration no one would ever use. Which rendered the measurement about as useless as could possibly be. About the only way to make it even more useless would be to use small files, enough ram to cache everything, and ….
No, I won’t name them. Just remember, as you read PR … there are PR numbers, and then there are end user tests on real code.
We occasionally lose business to folks with better PR numbers. Its interesting that the ratio of observed performance to PR performance is often very small.
The numbers we report … are the ones we measure. With realistic configs for end users. икони