Jamie Lucier (me@myisp.com) on 2/16/10 wrote:
---------------------------
>Paradox (paradox1000@gmail.com) on 2/16/10 wrote:
>---------------------------
>>Maybe the 250,000 is a mistake. According to the detailed report only 10,000 employees were used in the benchmark.
>>
>Agreed, upon closer look, I don't see 250,000 where I think it should appear in
>elsewhere in the detailed report...looks like shoddy auditing...which makes me question
>the validity of any of these benchmarks from Oracle.
>
>>>Also, I dug up the detailed report for these benchmarks from Oracle's site and
>>>the results seem questionable all around, but especially the POWER7 750 result appears
>>>to have a lot of guesstimates and "linear extrapolations" going on it in workload
>>>results table. This crude analysis on my part is based on the number of results
>>>that have what appears to be un-measured numbers. Heavy rounding going on whereas other results have none?
>>
>>No, the performance numbers are measured. The extrapolation is only for scaling
>>the numbers up to one hour. (I.e. if you can process 2 items in 1 minute, then you can proces 120 in an hour.)
>>
>>It is the total runtime that determines the benchmark performance. (This is in the minutes range for both machines.)
>>
>I think I understand that. But explain if you can chances you can get a result
>that ends in 5 or more zeros! It doesn't mention any rounding, it only mentions
>extrapolations which is quasi weasel words for guessing in my book.
>
>Workload POWER7 3.3GHz Xeon 5570 2.93 Ghz
>Payroll Processing 1,161,290 1,395,349
>Prepayments 4,500,000 1,052,632
>External Archive 423,529 468,750
>NACHA 18,000,000 20,000,000
>Checkwriter 4,500,000 2,727,273
>Costing 6,000,000 7,500,000


This is easily explained :-) For example the runtime for Checkwriter is 0.13 min in the table. But it is reasonable to assume the original time was measured in seconds, maybe not with too many decimals. 0.13 min is 7.8 seconds. If the correct time was 8 seconds, that would be 0.133333.. min., which is rounded to 0.13 in the table.

Now, 60/0.133333.. equals 450. Since the benchmark measures the time for 10,000 emplyees, the hourly rate is 450,000.

Of course you can argue this is a bit lax, but in reality it does not matter. The runtimes are very much dominated by "Payroll Processing" and "External Archive", and only the total runtime is used to get the benchmark number.



>
>Both servers appear to have questionable results. The difference is that results
>that appear to be actual measurements. The Xeon beats the POWER...not by much mind
>you. And with 8 physical cores instead of 6. But it appears that the POWER benchmark
>had SMT enabled and the Xeon didn't. All around these two benchmarks seems to be a mess.

The Xeon probably scored better with SMT disabled. You can also notice that the Power7 only used SMT2, not SMT4.