"F. Alfredo Rego" <[log in to unmask]> wrote:
>
>Our own Adager investigations have not yielded more than 5% to 10%
>improvement. And we could not get more concrete details from Robelle
>regarding the 35%.
>
>Nevertheless, assuming that a 5% to 10% is better than nothing, I have
>added the necessary logic to Adager to automatically enable prefetch for
>all Adager-transformed databases IF you are on MPE/iX 5.0 or later.
>(Naturally, as you would expect from Adager, I tell you that I have done so
>and WHY I have done so.)
>
>Before I release this automagic feature to the world, I would like to
>conduct a survey among this elite group of HP3000 users.
>
> - Would you like for Adager to automatically ENABLE
> prefetch for all Adager-transformed databases?
>
> - Would you NOT? (If so, why not?)
>
>I believe we should follow the recommendation reported by Robelle. But
>more conservative people might feel uneasy about it. Since I am more
>aggressive than conservative, I am asking for your feedback :-)
>
It has been a while since I have done any benchmarking, but as I recall from my
earlier tests and from discussions with IMAGE lab folks, this would not be a
good idea. I believe that the benefits accrue for fast machines with alot of
concurrent PUTs, DELETEs, and UPDATEs to specific data sets and little or no
memory pressure. The internal lock is held for a shorter time, permitting the
additional concurrency. For slower machines without this type of demand, no
benefit will be seen, or possibly even a degradation, since some data will
have to be fetched twice, if it gets swapped out before it is needed.
Unless Adager has a way of assessing this, I think you are better off leaving
things as they are.
For what it may be worth,
Steve
|