On Tue, 2011-10-18 at 04:40 +0100, Ben Hutchings wrote: > On Thu, 2011-10-13 at 07:05 +0100, Ian Campbell wrote: > > On Thu, 2011-10-13 at 03:27 +0100, Ben Hutchings wrote: > > > On Wed, 2011-10-12 at 14:58 +0100, Ian Campbell wrote: > > > > On Wed, 2011-10-12 at 14:11 +0100, Ben Hutchings wrote: > > > > > On Wed, 2011-10-12 at 08:26 +0100, Ian Campbell wrote: > > > > > > On Wed, 2011-10-12 at 08:46 +0300, Dmitry Musatov wrote: > > > > > > > The config option XEN_MAX_DOMAIN_MEMORY controls how much memory a > > > > > > > Xen instance is seeing. The default for 64bit is 32GB, which is the > > > > > > > reason that m2.4xlarge Amazon EC2 instances only report this amount of > > > > > > > memory. > > > > > > > Please set this limit to 70GB as there is a known restriction for > > > > > > > t1.micro instances at about 80GB. > > > > > > > Similar bug exists and Ubuntu where it's already fixed > > > > > > > (https://bugs.launchpad.net/ubuntu/+source/linux-ec2/+bug/667796) > > > > > > > > > > > > Is this the sort of change we can consider making in a stable update? > > > > > > I'm not at all sure, although my gut feeling is that it would be safe. > > > > > [...] > > > > > > > > > > I think so. But what is the trade-off? There must be some reason why > > > > > this isn't set to however many TB the kernel can support. > > > > > > > > It effects the amount of space set aside for the P2M table (the mapping > > > > of physical to machine addresses). In the kernel in Squeeze this space > > > > is statically reserved in BSS so increasing it will waste some more > > > > memory, according to the Kconfig comment it is 1 page per GB. > > > > > > > > In a more up to date kernel the space comes from BRK and is reclaimed if > > > > it is not used, MAX_DOMAIN_MEMORY was bumped to default to 128G in the > > > > same change. > > > > > > How intrusive is the change? Could we reasonably backport it? > > > > It was 58e05027b530 "xen: convert p2m to a 3 level tree" which I think > > is too big. IIRC there was a bunch of subsequent fixups to it as well, > > it was quite a subtle change. > > You didn't directly answer the questions, but that sounds like 'fairly' > and 'no'. Correct. Sorry, I thought "too big" covered both. > If I understand correctly, the memory cost of expanding the table to > cover 70GB is (70GB - 32GB) * 4KB / 1GB = 156KB. Is that right? Yes. > Since we don't have a specific flavour to support EC2, One option might be to increase the limit only for the xen flavour and leave the normal flavour where it is. That adds a rather unfortunate matrix to the selection of which flavour to use though, ideally I would prefer folks to be using the regular flavour in domU wherever possible. > and since some > people like to run domains with much less memory, I'm inclined to say > that this is 'wontfix' for squeeze. But I'm not sure just how small > they are likely to be (while still running Debian). Maybe the cost > isn't that significant. http://www.debian.org/releases/stable/amd64/ch03s04.html.en and http://www.debian.org/releases/stable/i386/ch03s04.html.en both say the minimum is 64M. We are talking about going from 128KB to 280KB reserved for the p2m. Which for a 64M machine is going from 0.20% to 0.43% of RAM overhead. I'm not sure if <64M is realistic. I have a (32-bit, physical) machine I use as a firewall which has 32M and apt-get and friends really do grind along (it's also an old Pentium with a tiny disk, so there are other factors in that). I think we are only talking about the limit for a 64 bit guest? I would guess that those are more unlikely to be given tiny amounts of RAM compared with 32 bit. Ian. -- Ian Campbell Current Noise: Monster Magnet - Bored With Sorcery Your picture of the world often changes just before you get it into focus.
Attachment:
signature.asc
Description: This is a digitally signed message part