Adam Litke <alitke at redhat.com> writes:>> On Tue Feb 11 06:01:10 UTC 2014, Rusty Russell wrote: >> Hi all! >> >> We're debating the design of the balloon for the OASIS spec. >> Noone likes the current one, but there are fundamental usage pattern >> questions which we're fumbling with. >> >> So if you know anyone who is using it in production? If, so, how? In >> particular, would you be happy with guests simply giving the host back >> whatever memory they can spare (as Xen's self-balloon does)? Or do >> you >> require the host-forcing approach? Comment or email please! > > Hi Rusty, > > I do not maintain any production setups but I have played with > ballooning (especially automatic ballooning) for quite some time now. > Most recently, I am working with the oVirt project [1] to enable > memory over-commitment and offer SLAs around VM memory usage.Hi Adam, Thanks for the comprehensive thoughts.> To address the question about whether the Xen self-balloon approach > would be enough... I think a guest-driven approach such as this would > work very well in self-hosted/private cloud deployments where a single > entity owns all of the virtual machines that are sharing memory. As > soon as you move to a "public" cloud environment where multiple > customers are sharing a single host then you will need a "bad cop" to > enforce some limits. (Yes I know ballooning always requires guest > cooperation, but when you combine it with punative cgroups on the host > the guest has a compelling reason to cooperate.) When I say "bad > cop", I mean a completely host-controlled balloon as we currently do > in oVirt with the Memory Overcommitment Manager [2]. This allows > customers to expect a certain minimum amount of performance.It's interesting that Dan Magenheimer made the opposite point: that if you're charging customers by the MB of memory, it's easy to get them to balloon themselves.> In order to support both modes of operation (at the same time) how > about supporting two virtio configuration variables in the balloon > driver: auto_min and auto_max. These variables would allow the host > to restrict the range in which the auto-balloon algorithm may operate. > Setting both to 0 would disable auto-ballooning and require all > inflate/deflate commands to come from the host. I think there are > some very interesting possibilities how auto-balloon can be combined > with host directed ballooning to yield good results in a variety of > configurations [3].I think we're headed to the same destination here; the variant which I came up with (and suggested to Daniel and Luiz, CC'd) is similar: the guest self-balloons, giving up pages when it can, but the host sets a ceiling. This way, if the host really needs to set a limit, it can: a disobedient guest will start paging. But generally, a guest should use its judgement to balloon its own pages as it can (below the ceiling). Thoughts? Rusty.
On 20/02/14 14:53 +1030, Rusty Russell wrote:>Adam Litke <alitke at redhat.com> writes: >>> On Tue Feb 11 06:01:10 UTC 2014, Rusty Russell wrote: >>> Hi all! >>> >>> We're debating the design of the balloon for the OASIS spec. >>> Noone likes the current one, but there are fundamental usage pattern >>> questions which we're fumbling with. >>> >>> So if you know anyone who is using it in production? If, so, how? In >>> particular, would you be happy with guests simply giving the host back >>> whatever memory they can spare (as Xen's self-balloon does)? Or do >>> you >>> require the host-forcing approach? Comment or email please! >> >> Hi Rusty, >> >> I do not maintain any production setups but I have played with >> ballooning (especially automatic ballooning) for quite some time now. >> Most recently, I am working with the oVirt project [1] to enable >> memory over-commitment and offer SLAs around VM memory usage. > >Hi Adam, > > Thanks for the comprehensive thoughts. > >> To address the question about whether the Xen self-balloon approach >> would be enough... I think a guest-driven approach such as this would >> work very well in self-hosted/private cloud deployments where a single >> entity owns all of the virtual machines that are sharing memory. As >> soon as you move to a "public" cloud environment where multiple >> customers are sharing a single host then you will need a "bad cop" to >> enforce some limits. (Yes I know ballooning always requires guest >> cooperation, but when you combine it with punative cgroups on the host >> the guest has a compelling reason to cooperate.) When I say "bad >> cop", I mean a completely host-controlled balloon as we currently do >> in oVirt with the Memory Overcommitment Manager [2]. This allows >> customers to expect a certain minimum amount of performance. > >It's interesting that Dan Magenheimer made the opposite point: that >if you're charging customers by the MB of memory, it's easy to get them >to balloon themselves.Sure, it's all about how the incentives are structured and what the workload is. Some people will insist on having a certain amount of memory "reserved" and available immediately. If you meter memory usage you would certainly shift the burden of conservation onto the guest and this could be preferred for some customers.> >> In order to support both modes of operation (at the same time) how >> about supporting two virtio configuration variables in the balloon >> driver: auto_min and auto_max. These variables would allow the host >> to restrict the range in which the auto-balloon algorithm may operate. >> Setting both to 0 would disable auto-ballooning and require all >> inflate/deflate commands to come from the host. I think there are >> some very interesting possibilities how auto-balloon can be combined >> with host directed ballooning to yield good results in a variety of >> configurations [3]. > >I think we're headed to the same destination here; the variant which I >came up with (and suggested to Daniel and Luiz, CC'd) is similar: the >guest self-balloons, giving up pages when it can, but the host sets a >ceiling. > >This way, if the host really needs to set a limit, it can: a disobedient >guest will start paging. But generally, a guest should use its >judgement to balloon its own pages as it can (below the ceiling).It sounds similar but it sounds like you are suggesting one limit value and I am suggesting two. Your ceiling value sounds like a soft limit on total guest memory (aka minimum balloon size). This is the more important limit of the two I have suggested. Do you think it's also worthwhile to have a maximum balloon size (floor value) to keep the allowable balloon size between two points? -- Adam Litke
On Thu, 20 Feb 2014 08:17:06 -0500 Adam Litke <alitke at redhat.com> wrote:> On 20/02/14 14:53 +1030, Rusty Russell wrote: > >Adam Litke <alitke at redhat.com> writes: > >>> On Tue Feb 11 06:01:10 UTC 2014, Rusty Russell wrote: > >>> Hi all! > >>> > >>> We're debating the design of the balloon for the OASIS spec. > >>> Noone likes the current one, but there are fundamental usage pattern > >>> questions which we're fumbling with. > >>> > >>> So if you know anyone who is using it in production? If, so, how? In > >>> particular, would you be happy with guests simply giving the host back > >>> whatever memory they can spare (as Xen's self-balloon does)? Or do > >>> you > >>> require the host-forcing approach? Comment or email please! > >> > >> Hi Rusty, > >> > >> I do not maintain any production setups but I have played with > >> ballooning (especially automatic ballooning) for quite some time now. > >> Most recently, I am working with the oVirt project [1] to enable > >> memory over-commitment and offer SLAs around VM memory usage. > > > >Hi Adam, > > > > Thanks for the comprehensive thoughts. > > > >> To address the question about whether the Xen self-balloon approach > >> would be enough... I think a guest-driven approach such as this would > >> work very well in self-hosted/private cloud deployments where a single > >> entity owns all of the virtual machines that are sharing memory. As > >> soon as you move to a "public" cloud environment where multiple > >> customers are sharing a single host then you will need a "bad cop" to > >> enforce some limits. (Yes I know ballooning always requires guest > >> cooperation, but when you combine it with punative cgroups on the host > >> the guest has a compelling reason to cooperate.) When I say "bad > >> cop", I mean a completely host-controlled balloon as we currently do > >> in oVirt with the Memory Overcommitment Manager [2]. This allows > >> customers to expect a certain minimum amount of performance. > > > >It's interesting that Dan Magenheimer made the opposite point: that > >if you're charging customers by the MB of memory, it's easy to get them > >to balloon themselves. > > Sure, it's all about how the incentives are structured and what the > workload is. Some people will insist on having a certain amount of > memory "reserved" and available immediately. If you meter memory > usage you would certainly shift the burden of conservation onto the > guest and this could be preferred for some customers. > > > > >> In order to support both modes of operation (at the same time) how > >> about supporting two virtio configuration variables in the balloon > >> driver: auto_min and auto_max. These variables would allow the host > >> to restrict the range in which the auto-balloon algorithm may operate. > >> Setting both to 0 would disable auto-ballooning and require all > >> inflate/deflate commands to come from the host. I think there are > >> some very interesting possibilities how auto-balloon can be combined > >> with host directed ballooning to yield good results in a variety of > >> configurations [3]. > > > >I think we're headed to the same destination here; the variant which I > >came up with (and suggested to Daniel and Luiz, CC'd) is similar: the > >guest self-balloons, giving up pages when it can, but the host sets a > >ceiling. > > > >This way, if the host really needs to set a limit, it can: a disobedient > >guest will start paging. But generally, a guest should use its > >judgement to balloon its own pages as it can (below the ceiling). > > It sounds similar but it sounds like you are suggesting one limit > value and I am suggesting two. Your ceiling value sounds like a soft > limit on total guest memory (aka minimum balloon size). This is the > more important limit of the two I have suggested. Do you think it's > also worthwhile to have a maximum balloon size (floor value) to keep > the allowable balloon size between two points?I was already planning for everything you asked (QMP commands to disable/enable automatic ballooning, min and max sizes). I still have to think a bit how those settings will fit in the guest-led design, but it it should be fine in principle.
Adam Litke <alitke at redhat.com> writes:> On 20/02/14 14:53 +1030, Rusty Russell wrote: >>I think we're headed to the same destination here; the variant which I >>came up with (and suggested to Daniel and Luiz, CC'd) is similar: the >>guest self-balloons, giving up pages when it can, but the host sets a >>ceiling. >> >>This way, if the host really needs to set a limit, it can: a disobedient >>guest will start paging. But generally, a guest should use its >>judgement to balloon its own pages as it can (below the ceiling). > > It sounds similar but it sounds like you are suggesting one limit > value and I am suggesting two. Your ceiling value sounds like a soft > limit on total guest memory (aka minimum balloon size). This is the > more important limit of the two I have suggested. Do you think it's > also worthwhile to have a maximum balloon size (floor value) to keep > the allowable balloon size between two points?It's a little simpler to have a ceiling only. And if everyone (guests and host) are feeling like they have plenty of memory, it's probably best sitting in the host. 1) It can quickly go to a guest if necessary. 2) The host could coalesce/shuffle memory. 3) The host could turn off RAM to save power. Cheers, Rusty.