Open Stack

Shihazhang,

     I really believe we need the RPC refactor done for this cycle,
and given the close deadlines we have (July 10 for spec submission and 
July 20 for spec approval).

     Don't you think it's going to be better to split the work in several
specs?

1) ipset optimization   (you)
2) sg rpc optimization (without fanout) (me)
3) sg rpc optimization (with fanout) (edouard, you , me)


    This way we increase the chances of having part of this for the Juno
cycle. If we go for something too complicated is going to take more time 
for approval.


   Also, I proposed the details of "2", trying to bring awareness on the
topic, as I have been working with the scale lab in Red Hat to find
and understand those issues, I have a very good knowledge of the problem
and I believe I could make a very fast advance on the issue at the
RPC level.

   Given that, I'd like to work on this specific part, whether or
not we split the specs, as it's something we believe critical for
neutron scalability and thus, *nova parity*.

    I will start a separate spec for "2", later on, if you find it ok,
we keep them as separate ones, if you believe having just 1 spec (for 1
  & 2) is going be safer for juno-* approval, then we can incorporate my 
spec in yours, but then "add-ipset-to-security" is not a good spec title 
to put all this together.


    Best regards,
Miguel Ángel.


On 07/02/2014 03:37 AM, shihanzhang wrote:
>
> hi Miguel Angel Ajo Pelayo!
> I agree with you and modify my spes, but I will also optimization the
> RPC from security group agent to neutron server.
> Now the modle is 'port[rule1,rule2...], port...', I will change it to
> 'port[sg1, sg2..]', this can reduce the size of RPC respose message from
> neutron server to security group agent.
>
> At 2014-07-01 09:06:17, "Miguel Angel Ajo Pelayo" <mangelajo at redhat.com> wrote:
>>
>>
>>Ok, I was talking with Édouard @ IRC, and as I have time to work
>>into this problem, I could file an specific spec for the security
>>group RPC optimization, a masterplan in two steps:
>>
>>1) Refactor the current RPC communication for security_groups_for_devices,
>>   which could be used for full syncs, etc..
>>
>>2) Benchmark && make use of a fanout queue per security group to make
>>   sure only the hosts with instances on a certain security group get
>>   the updates as they happen.
>>
>>@shihanzhang do you find it reasonable?
>>
>>
>>
>>----- Original Message -----
>>> ----- Original Message -----
>>> > @Nachi: Yes that could a good improvement to factorize the RPC mechanism.
>>> >
>>> > Another idea:
>>> > What about creating a RPC topic per security group (quid of the RPC topic
>>> > scalability) on which an agent subscribes if one of its ports is associated
>>> > to the security group?
>>> >
>>> > Regards,
>>> > Édouard.
>>> >
>>> >
>>>
>>>
>>> Hmm, Interesting,
>>>
>>> @Nachi, I'm not sure I fully understood:
>>>
>>>
>>> SG_LIST [ SG1, SG2]
>>> SG_RULE_LIST = [SG_Rule1, SG_Rule2] ..
>>> port[SG_ID1, SG_ID2], port2 , port3
>>>
>>>
>>> Probably we may need to include also the
>>> SG_IP_LIST = [SG_IP1, SG_IP2] ...
>>>
>>>
>>> and let the agent do all the combination work.
>>>
>>> Something like this could make sense?
>>>
>>> Security_Groups = {SG1:{IPs:[....],RULES:[....],
>>>                    SG2:{IPs:[....],RULES:[....]}
>>>                   }
>>>
>>> Ports = {Port1:[SG1, SG2], Port2: [SG1] .... }
>>>
>>>
>>> @Edouard, actually I like the idea of having the agent subscribed
>>> to security groups they have ports on... That would remove the need to
>>> include
>>> all the security groups information on every call...
>>>
>>> But would need another call to get the full information of a set of security
>>> groups
>>> at start/resync if we don't already have any.
>>>
>>>
>>> >
>>> > On Fri, Jun 20, 2014 at 4:04 AM, shihanzhang < ayshihanzhang at 126.com >
>>> > wrote:
>>> >
>>> >
>>> >
>>> > hi Miguel Ángel,
>>> > I am very agree with you about the following point:
>>> > >  * physical implementation on the hosts (ipsets, nftables, ... )
>>> > --this can reduce the load of compute node.
>>> > >  * rpc communication mechanisms.
>>> > -- this can reduce the load of neutron server
>>> > can you help me to review my BP specs?
>>> >
>>> >
>>> >
>>> >
>>> >
>>> >
>>> >
>>> > At 2014-06-19 10:11:34, "Miguel Angel Ajo Pelayo" < mangelajo at redhat.com >
>>> > wrote:
>>> > >
>>> > >  Hi it's a very interesting topic, I was getting ready to raise
>>> > >the same concerns about our security groups implementation, shihanzhang
>>> > >thank you for starting this topic.
>>> > >
>>> > >  Not only at low level where (with our default security group
>>> > >rules -allow all incoming from 'default' sg- the iptable rules
>>> > >will grow in ~X^2 for a tenant, and, the
>>> > >"security_group_rules_for_devices"
>>> > >rpc call from ovs-agent to neutron-server grows to message sizes of
>>> > >>100MB,
>>> > >generating serious scalability issues or timeouts/retries that
>>> > >totally break neutron service.
>>> > >
>>> > >   (example trace of that RPC call with a few instances
>>> > > http://www.fpaste.org/104401/14008522/ )
>>> > >
>>> > >  I believe that we also need to review the RPC calling mechanism
>>> > >for the OVS agent here, there are several possible approaches to breaking
>>> > >down (or/and CIDR compressing) the information we return via this api
>>> > >call.
>>> > >
>>> > >
>>> > >   So we have to look at two things here:
>>> > >
>>> > >  * physical implementation on the hosts (ipsets, nftables, ... )
>>> > >  * rpc communication mechanisms.
>>> > >
>>> > >   Best regards,
>>> > >Miguel Ángel.
>>> > >
>>> > >----- Mensaje original -----
>>> > >
>>> > >> Do you though about nftables that will replace {ip,ip6,arp,eb}tables?
>>> > >> It also based on the rule set mechanism.
>>> > >> The issue in that proposition, it's only stable since the begin of the
>>> > >> year
>>> > >> and on Linux kernel 3.13.
>>> > >> But there lot of pros I don't list here (leverage iptables limitation,
>>> > >> efficient update rule, rule set, standardization of netfilter
>>> > >> commands...).
>>> > >
>>> > >> Édouard.
>>> > >
>>> > >> On Thu, Jun 19, 2014 at 8:25 AM, henry hly < henry4hly at gmail.com >
>>> > >> wrote:
>>> > >
>>> > >> > we have done some tests, but have different result: the performance is
>>> > >> > nearly
>>> > >> > the same for empty and 5k rules in iptable, but huge gap between
>>> > >> > enable/disable iptable hook on linux bridge
>>> > >>
>>> > >
>>> > >> > On Thu, Jun 19, 2014 at 11:21 AM, shihanzhang < ayshihanzhang at 126.com
>>> > >> > >
>>> > >> > wrote:
>>> > >>
>>> > >
>>> > >> > > Now I have not get accurate test data, but I can confirm the
>>> > >> > > following
>>> > >> > > points：
>>> > >> >
>>> > >>
>>> > >> > > 1. In compute node, the iptable's chain of a VM is liner, iptable
>>> > >> > > filter
>>> > >> > > it
>>> > >> > > one by one, if a VM in default security group and this default
>>> > >> > > security
>>> > >> > > group have many members, but ipset chain is set, the time ipset
>>> > >> > > filter
>>> > >> > > one
>>> > >> > > and many member is not much difference.
>>> > >> >
>>> > >>
>>> > >> > > 2. when the iptable rule is very large, the probability of failure
>>> > >> > > that
>>> > >> > > iptable-save save the iptable rule is very large.
>>> > >> >
>>> > >>
>>> > >
>>> > >> > > At 2014-06-19 10:55:56, "Kevin Benton" < blak111 at gmail.com > wrote:
>>> > >> >
>>> > >>
>>> > >
>>> > >> > > > This sounds like a good idea to handle some of the performance
>>> > >> > > > issues
>>> > >> > > > until
>>> > >> > > > the ovs firewall can be implemented down the the line.
>>> > >> > >
>>> > >> >
>>> > >>
>>> > >> > > > Do you have any performance comparisons?
>>> > >> > >
>>> > >> >
>>> > >>
>>> > >> > > > On Jun 18, 2014 7:46 PM, "shihanzhang" < ayshihanzhang at 126.com >
>>> > >> > > > wrote:
>>> > >> > >
>>> > >> >
>>> > >>
>>> > >
>>> > >> > > > > Hello all,
>>> > >> > > >
>>> > >> > >
>>> > >> >
>>> > >>
>>> > >
>>> > >> > > > > Now in neutron, it use iptable implementing security group, but
>>> > >> > > > > the
>>> > >> > > > > performance of this implementation is very poor, there is a bug:
>>> > >> > > > > https://bugs.launchpad.net/neutron/+bug/1302272 to reflect this
>>> > >> > > > > problem.
>>> > >> > > > > In
>>> > >> > > > > his test, w ith default security groups(which has remote
>>> > >> > > > > security
>>> > >> > > > > group),
>>> > >> > > > > beyond 250-300 VMs, there were around 6k Iptable rules on evry
>>> > >> > > > > compute
>>> > >> > > > > node,
>>> > >> > > > > although his patch can reduce the processing time, but it don't
>>> > >> > > > > solve
>>> > >> > > > > this
>>> > >> > > > > problem fundamentally. I have commit a BP to solve this problem:
>>> > >> > > > > https://blueprints.launchpad.net/neutron/+spec/add-ipset-to-security
>>> > >> > > > > >> > > >
>>> > >> > >
>>> > >> >
>>> > >>
>>> > >> > > > > There are other people interested in this it?
>>> > >> > > >
>>> > >> > >
>>> > >> >
>>> > >>
>>> > >
>>> > >> > > > > _______________________________________________
>>> > >> > > >
>>> > >> > >
>>> > >> >
>>> > >>
>>> > >> > > > > OpenStack-dev mailing list
>>> > >> > > >
>>> > >> > >
>>> > >> >
>>> > >>
>>> > >> > > > > OpenStack-dev at lists.openstack.org >> > > >
>>> > >> > >
>>> > >> >
>>> > >>
>>> > >> > > > > http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>>> > >> > > > > >> > > >
>>> > >> > >
>>> > >> >
>>> > >>
>>> > >
>>> > >> > > _______________________________________________
>>> > >> >
>>> > >>
>>> > >> > > OpenStack-dev mailing list
>>> > >> >
>>> > >>
>>> > >> > > OpenStack-dev at lists.openstack.org >> >
>>> > >>
>>> > >> > > http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev >>
>>> > >> > > >
>>> > >>
>>> > >
>>> > >> > _______________________________________________
>>> > >>
>>> > >> > OpenStack-dev mailing list
>>> > >>
>>> > >> > OpenStack-dev at lists.openstack.org >>
>>> > >> > http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev >>
>>> > >
>>> > >> _______________________________________________
>>> > >> OpenStack-dev mailing list
>>> > >> OpenStack-dev at lists.openstack.org >>
>>> > >> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev >
>>> > >_______________________________________________
>>> > >OpenStack-dev mailing list
>>> > > OpenStack-dev at lists.openstack.org >
>>> > > http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>>> >
>>> >
>>> >
>>> > _______________________________________________
>>> > OpenStack-dev mailing list
>>> > OpenStack-dev at lists.openstack.org
>>> > http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>>> >
>>> >
>>> >
>>> > _______________________________________________
>>> > OpenStack-dev mailing list
>>> > OpenStack-dev at lists.openstack.org
>>> > http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>>> >
>>>
>>> _______________________________________________
>>> OpenStack-dev mailing list
>>> OpenStack-dev at lists.openstack.org
>>> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>>>
>>
>>_______________________________________________
>>OpenStack-dev mailing list
>>OpenStack-dev at lists.openstack.org
>>http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>
>
>
>
>
> _______________________________________________
> OpenStack-dev mailing list
> OpenStack-dev at lists.openstack.org
> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>