On Fri, Jan 31, 2025 at 06:46:21AM +0100, Stefano Brivio wrote: > On Thu, 30 Jan 2025 19:54:17 +1100 > David Gibson wrote: > > > On Thu, Jan 30, 2025 at 09:32:36AM +0100, Stefano Brivio wrote: > > > On Thu, 30 Jan 2025 18:38:22 +1100 > > > David Gibson wrote: > > > > > > > Right, but in the present draft you pay that cost whether or not > > > > you're actually using the flows. Unfortunately a busy server with > > > > heaps of active connections is exactly the case that's likely to be > > > > most sensitve to additional downtime, but there's not really any > > > > getting around that. A machine with a lot of state will need either > > > > high downtime or high migration bandwidth. > > > > > > It's... sixteen megabytes. A KubeVirt node is only allowed to perform up > > > to _four_ migrations in parallel, and that's our main use case at the > > > moment. "High downtime" is kind of relative. > > > > Certainly. But I believe it's typical to aim for downtimes in the > > ~100ms range. > > > > > > But, I'm really hoping we can move relatively quickly to a model where > > > > a guest with only a handful of connections _doesn't_ have to pay that > > > > 128k flow cost - and can consequently migrate ok even with quite > > > > constrained migration bandwidth. In that scenario the size of the > > > > header could become significant. > > > > > > I think the biggest cost of the full flow table transfer is rather code > > > that's a bit quicker to write (I just managed to properly set sequences > > > on the target, connections don't quite "flow" yet) but relatively high > > > maintenance (as you mentioned, we need to be careful about every single > > > field) and easy to break. > > > > Right. And with this draft we can't even change the size of the flow > > table without breaking migration. That seems like a thing we might > > well want to change. > > Why? The size of the flow table hasn't changed since it was added. I Which wasn't _that_ long ago. It just seems like a really obvious constant to tune to me, and one which it would be surprising if it broken migration. > don't see a reason to improve this if we don't want to transfer the > flow table anyway. I don't follow. Do you mean not transferring the hash table? This is not relevant to that, I'm talking about the size of the base flow table, not the hash table. Or do you mean not transferring the flow table as a whole, but rather entry by entry? In that case I'm seeing it as exactly the mechanism to improve this. > > > I would like to quickly complete the whole flow first, because I think > > > we can inform design and implementation decisions much better at that > > > point, and we can be sure it's feasible, > > > > That's fair. > > > > > but I'm not particularly keen > > > to merge this patch like it is, if we can switch it relatively swiftly > > > to an implementation where we model a smaller fixed-endian structure > > > with just the stuff we need. > > > > So, there are kind of two parts to this: > > > > 1) Only transferring active flow entries, and not transferring the > > hash table > > > > I think this is pretty easy. It could be done with or without > > preserving flow indicies. Preserving makes for debug log continuity > > between the ends, but not preserving lets us change the size of the > > flow table without breaking migration. > > I would just add prints on migration showing how old flow indices map > to new ones. That's possible, although it would mean transferring the old indices, which is not otherwise strictly necessary. What we could do easily is a debug log similar to the "new flow" logs but for "immigrated flow". > > 2) Only transferring the necessary pieces of each entry, and using a > > fixed representation of each piece > > > > This is harder. Not *super* hard, I think, but definitely trickier > > than (1) > > > > > And again, to be a bit more sure of which stuff we need in it, the full > > > flow is useful to have implemented. > > > > > > Actually the biggest complications I see in switching to that approach, > > > from the current point, are that we need to, I guess: > > > > > > 1. model arrays (not really complicated by itself) > > > > So here, I actually think this is simpler if we don't attempt to have > > a declarative approach to defining the protocol, but just write > > functions to implement it. > > > > > 2. have a temporary structure where we store flows instead of using the > > > flow table directly (meaning that the "data model" needs to logically > > > decouple source and destination of the copy) > > > > Right.. I'd really prefer to "stream" in the entries one by one, > > rather than having a big staging area. That's even harder to do > > declaratively, but I think the other advantages are worth it. > > > > > 3. batch stuff to some extent. We'll call socket() and connect() once > > > for each socket anyway, obviously, but sending one message to the > > > TCP_REPAIR helper for each socket looks like a rather substantial > > > and avoidable overhead > > > > I don't think this actually has a lot of bearing on the protocol. I'd > > envisage migrate_target() decodes all the information into the > > target's flow table, then migrate_target_post() steps through all the > > flows re-establishing the connections. Since we've already parsed the > > protocol at that point, we can make multiple passes: one to gather > > batches and set TCP_REPAIR, another through each entry to set the > > values, and a final one to clear TCP_REPAIR in batches. > > Ah, right, I didn't think of using the target flow table directly. That > has the advantage that the current code I'm writing to reactivate flows > from the flow table can be recycled as it is. Possibly - it might need to do some slightly different things: regenerating some fields from redundant data maybe, and/or re-hashing the entries. But certainly the structure should be similar, yes. > > > > > > It's both easier to do > > > > > > and a bigger win in most cases. That would dramatically reduce the > > > > > > size sent here. > > > > > > > > > > Yep, feel free. > > > > > > > > It's on my queue for the next few days. > > > > > > To me this part actually looks like the biggest priority after/while > > > getting the whole thing to work, because we can start right with a 'v1' > > > which looks more sustainable. > > > > > > And I would just get stuff working on x86_64 in that case, without even > > > implementing conversions and endianness switches etc. > > > > Right. Given the number of options here, I think it would be safest > > to go in expecting to go through a few throwaway protocol versions > > before reaching one we're happy enough to support long term. > > > > To ease that process, I'm wondering if we should, add a default-off > > command line option to enable migration. For now, enabling it would > > print some sort of "migration is experimental!" warning. Once we have > > a stream format we're ok with, we can flip it to on-by-default, but we > > don't maintain receive compatibility for the experimental versions > > leading up to that. > > It looks like unnecessary code churn to me. It doesn't need to be > merged if it's work in progress. You can also push stuff to a temporary > branch if needed. Eh, just thought merging might save us some rebase work against any other pressing changes we need. > It can also be merged and not documented for a while, as long as it > doesn't break existing functionality. I'd be a bit cautious about this. AIUI, right now if you attempt to migrate, qemu will simply fail it because we don't respond to the migration commands. Having enough merged that qemu won't outright fail the migration, but it won't reliably work seems like a bad idea. -- David Gibson (he or they) | I'll have my music baroque, and my code david AT gibson.dropbear.id.au | minimalist, thank you, not the other way | around. http://www.ozlabs.org/~dgibson