fivemack: (Default)
Tom Womack ([personal profile] fivemack) wrote2006-07-21 01:38 pm
Entry tags:

Why demonstrate 3 million TPC-C?

There's a transaction-processing benchmark, results listed at www.tpc.org, for which major computer manufacturers are prepared to spend millions of dollars of engineering time and use tens of millions of dollars worth of hardware.

The current top entries are offering rates of a couple of million transactions per minute, which translates to between one and two trillion transactions a year since there are almost exactly half a million minutes in a year.

I've just looked at ebay's financial statements, which indicate that 2.5 billion items are sold through ebay annually; if we assume that each bid is a transaction and that each item gets twenty bids, that's a hundred thousand transactions a minute.

Tesco's sales are on the order of £40 billion a year; even if we assume that each item on a bill is a transaction, the average Tesco item cost more than 50p, so that's less than 80 billion transactions a year, 160,000 a minute. Wal-Mart has about five times the sales of Tesco, which brings you to half a trillion; a million a minute.

What exactly is the point to IBM or to HP of demonstrating a single machine capable of handling all the transactions at every supermarket in the European Union? This seems the kind of machine of which they can sell one.

[identity profile] mobbsy.livejournal.com 2006-07-21 01:07 pm (UTC)(link)
Not all transactions are created equal - and the TCP-C ones come under some criticism for being unrealistically simple and lightweight for modern transaction processing loads. The benchmark was initally created in 1992, and while it's been updated since is still essentially the same workload.

Like a lot of other aging benchmarks, the exact numbers become meaningless, but it is still useful as a guide for comparing OLTP performance between vendors. That said, it's now so old that even that purpose is quite eroded, but as yet there's nothing really better (excluding possibly some of the SAP benchmarks, but they've the problem of being tied to a single software vendor).

http://www.tpc.org/tpce/tpc-e.asp is the discussion document on a suggested replacement for TPC-C

[identity profile] randwolf.livejournal.com 2006-07-21 06:16 pm (UTC)(link)
I think the target market is banks and the like, not supermarkets. The NSA might also be interested.

[identity profile] mobbsy.livejournal.com 2006-07-21 11:11 pm (UTC)(link)
Not at all. I've recently used the TPC-C benchmarks extensively to inform my choice of the correct server for my employer (an online retailer), and while not top-end (and certainly not equipped with the ludicrous 6400 disks and 2 TB RAM of the IBM 64-way p595 TPC-C benchmark), we ended up with a fairly hefty system.

In a previous job, I've benchmarked on a 32-way p590. That was targeted at telco billing, and one such server would be a realistic platform for a medium to large mobile telco operating in a country like UK, Germany or France.

Another way to look at it is that the current dual-core quad Opteron systems are now putting in TPC-C results comperable to the top-end results of 6 years ago (e.g. 32-way Alpha 21264). What people do with the capacity expands to meet what's available.

[identity profile] born2benchmark.livejournal.com 2006-09-28 07:55 pm (UTC)(link)
I have tried to create an alternative to the TPC transaction benchmarks.
Information on the 'Small Business Transaction Benchmark(TM)' can be found at http://www.worlds-fastest.com. Click on the link 'Transaction Benchmark Tests'. My 'standard configuration' tests place limits on the machine configuration with identical O/S, tuning, application software, memory size and disk drives.