LWN: Comments on "Virtual Memory II: the return of objrmap"

Virtual Memory II: the return of objrmap

khim — 2004-05-24T07:19:50+00:00

Low memory on 32bit systems is still 2Gb! And if it's not enough for some structures on 32Gb system then obviously something is wrong: 10% for book-keeping is too much IMO (cache misses and all). So this patch is sane. True, only huge 32bit systems make it 2.6 and not 2.7 material but patch itself is sane - it's good for huge 64bit systems as well (not sure about small systems), just not essential there.

Virtual Memory II: the return of objrmap

leandro — 2004-04-16T18:20:34+00:00

> what "Intel killed the Alpha" means

Alpha ceased to be developed.

> Alphas are available now (e.g., Compaq ES-47)

Expensive, limited, substandard. In other words, not developed neither as technology nor as a platform.

> the Alpha technology is planned to be integrated into the Itanium

Processor architecture is not like Lego where you can mix and match. The Itanium and Alpha architectures are fundamentally different and philosophically opposed. Some Alpha tricks may be incorporated into Itanium, but it will never see the potential Alpha had, and POWER still has but with a different focus. Some argue that nothing has the potential Alpha had.

Virtual Memory II: the return of objrmap

alpharomeo — 2004-03-21T15:34:27+00:00

Not sure what "Intel killed the Alpha" means. Alphas are available now (e.g., Compaq ES-47) and the Alpha technology is planned to be integrated into the Itanium product line starting in '06. Do you know something different?

Virtual Memory II: the return of objrmap

alpharomeo — 2004-03-21T15:31:29+00:00

Page clustering seems like the obvious solution. From one point of view, it is the 4K page size that is obsolete, not the 32-bit addressing. When can we anticipate having a page clustering option available?

Virtual Memory II: the return of objrmap

leandro — 2004-03-18T13:42:06+00:00

I guess even Linus sometimes bows to pressure. After all, all this complication is quite unnecessary, it is a decade now that we've had 64 bits processors. Nothing but Wintel FUD and proprietary software prevents users from running 64 bits now.

It could be argued that 64 bits vendors haven't been doing the right thing. Now Linus is on a POWER64 machine, but he should have been there for a long time, or on UltraSPARC. A pity Intel killed the Alpha, which was once Linus' platform. Also other developers should have been long ago given such systems.

I hope the BSDs and the Hurd stick to sanity.

Virtual Memory II: the return of objrmap

riel — 2004-03-11T18:51:07+00:00

There is some preview code available that would decrease the CPU overhead of objrmap, probably to acceptable levels. If that code works as well as it's supposed to work (it's a new data structure, not yet well tested, etc...) there's a reasonable chance the pte based rmap can be replaced.

It's good to see that the last reason that caused me to make the pte based rmap code is finally dissolving. A well working object based rmap is much lighter weight...