[pve-devel] [PATCH v2 qemu-server 00/16] migration: split phases

[Alexandre DERUMIER]

>>A loop where SSL etc. was done on the same CPU got us ~1 GBps here, cross server
>>it could be faster - we need to still evaluate that. But for initial PoC it seemed
>>to be enough - we had websocket in http-server already.


1GBps or 1Gbps (100MB/s) ?  because 1Gbps is really too slow for me
(with fast ssd/nvme storage I'm not sure about this convergence at the end of block mirroring)

also, reusing 8006, the kind of certificat used (ecc / rsa) + cipher could make big differences on encryption
speed.


>>But, having a "insecure mode" to satisfy some ( ;-P ) setups was also already
>>discussed here.

Yep, make my 40Gbps nic happy ;)


I'll do tests and send results

----- Mail original -----
De: "Thomas Lamprecht" <t.lamprecht at proxmox.com>
À: "pve-devel" <pve-devel at pve.proxmox.com>, "Alexandre Derumier" <aderumier at odiso.com>
Envoyé: Lundi 9 Mars 2020 18:36:44
Objet: Re: [pve-devel] [PATCH v2 qemu-server 00/16] migration: split phases

On 3/9/20 5:43 PM, Alexandre DERUMIER wrote: 
>>> Fabian is also looking into external migrate, posted a initial proof-of-concept 
>>> last week[0]. The currently planned design is roughly: 
>>> * having "remotes", which can be any nodes, or a whole cluster 
>>> * we want to use the new API Token mechanisms for remotes - this allows 
>>> nice integration where one can also revoke access from another node to 
>>> theirs. 
>>> * migration tunnel will for use WebSocket for secure data transfer, this 
>>> allows to do a HTTPS request over API then switch protocol (i.e., what 
>>> HTTP/2 would get us with multiplexing but not as nice) 
> Oh, that's great ! I totally miss it . I'm curious to see performance of websocket for such big data transfert. 
> 

A loop where SSL etc. was done on the same CPU got us ~1 GBps here, cross server 
it could be faster - we need to still evaluate that. But for initial PoC it seemed 
to be enough - we had websocket in http-server already. 

But, having a "insecure mode" to satisfy some ( ;-P ) setups was also already 
discussed here. 

> (BTW, I'm still dreaming of a multi-cluster pve-manager ;) 
> 

We too, this would be the second bigger step towards that; the first one was API 
tokens :) 

> 
>>> We may naturally still want to see if some patches of the split up make 
>>> sense to apply now. 
>>> Just that you know, not that you invest too much time in this now :-) 
> ok, thanks. (I had a need a work for this week, that's why I hve sent it) 
> 

OK 

> I'll test the POC tomorrow. 

great, but as you can read from the commit message it's really early, more 
for us to decide if this goes the right direction before investing days to 
weeks of polishing work :)