On Thu, 31 May 2007 23:36:32 +0200, PFC <lists@peufeu.com> wrote:
> On Thu, 31 May 2007 22:20:09 +0200, Vivek Khera <vivek@khera.org> wrote:
>
>>
>> On May 25, 2007, at 5:28 PM, Tom Lane wrote:
>>
>>> That's true at the level of DDL operations, but AFAIK we could
>>> parallelize table-loading and index-creation steps pretty effectively
>>> --- and that's where all the time goes.
>>
>> I would be happy with parallel builds of the indexes of a given table.
>> That way you have just one scan of the whole table to build all its
>> indexes.
Just did a test :
- large table (does not fit in RAM)
- rows with text column (forum posts)
- about 700K rows
Time to create 3 indexes : 61 s
Time to create 3 indexes with 3 simultaneous connections : 22 s
That's what you would expect...
vmstat shows the data is really loaded from disk, once with the 3
threads, 3 times when indexes are created one at a time.