Hi Andrew,
I am working on ACL plugin SF+SL optimization on ARM servers.
I am finding prefetches in ACL node is becoming bottle neck. I see
performance improvements on both SL & SF mode, when SF mode bihash table
related prefetching is disabled.
I need some help with right ACL config to verify my patch.
I did the testing with Ingress ACL -- 1 Rule and 50 Rules (Rule: <SIP, DIP,
UDP, SPORT, DPORT> - DPORT is incremented). The Traffic match all the 50 rules.
When I tried to add 100 rules on the same rule set in SF mode:
"acl_add_replace -1 ipv4 permit+reflect src 192.81.1.1/32 dst 192.82.1.1/32
proto 17 sport 100 dport 1,
... ,
ipv4 permit+reflect src 192.81.1.1/32 dst 192.82.1.1/32 proto 17 sport 100
dport 100",
I see only 48 rules in show tables and 48th rule is added as “permit” all
and not “permit + reflect”. Does it mean <0 – 47> rules will be SF and the rest
will be in SL mode ?
"
vpp# show acl-plugin acl
acl-index 0 count 49 tag {}
0: ipv4 permit+reflect src 192.81.1.1/32 dst 192.82.1.1/32 proto 17
sport 100 dport 1
....
47: ipv4 permit+reflect src 192.81.1.1/32 dst 192.82.1.1/32 proto 17
sport 100 dport 48
48: ipv4 permit src 0.0.0.0/0 dst 0.0.0.0/0 proto 0 sport 0-65535 dport
0-65535
applied inbound on sw_if_index: 1
used in lookup context index: 0
"
1. Is there a limit of 48 on number of rules that can be added into the Rule
table (acl-index 0) in SF mode ?
2. Whether 48 rules in a ruleset is good enough to verify my optimization
patch (Traffic flow will match all the 48 rules) ?
3. Can I associate more than 1 ACL rule set to an ingress interface (like
“vat# acl_interface_set_acl_list TenGigabitEthernet1/0/0 input 0 1 2”) ? Each
Rule set 0, 1, 2 will have different ACL rules. Do I need to test this case
also to study the performance gain ?
4. In SL mode, When I tried to add 100 rules, only 53 rules are seen in show
table. 53rd rule is added as permit all (Should I read it as permit all ?). Is
there a limit on number of rules in SL mode ?
“
vpp# show acl-plugin acl
acl-index 0 count 54 tag {}
0: ipv4 permit src 192.81.1.1/32 dst 192.82.1.1/32 proto 17 sport 100
dport 1
….
52: ipv4 permit src 192.81.1.1/32 dst 192.82.1.1/32 proto 17 sport 100
dport 53
53: ipv4 permit src 0.0.0.0/0 dst 0.0.0.0/0 proto 0 sport 0-65535
dport 0-65535
applied inbound on sw_if_index: 1
used in lookup context index: 0
“
Thanks
Govind
> -----Original Message-----
> From: vpp-dev@lists.fd.io <vpp-dev@lists.fd.io> On Behalf Of Govindarajan
> Mohandoss via Lists.Fd.Io
> Sent: Friday, March 27, 2020 11:32 AM
> To: Andrew 👽 Yourtchenko <ayour...@gmail.com>
> Cc: vpp-dev@lists.fd.io
> Subject: Re: [vpp-dev] ACL question
>
> Thank you very much Andrew !! I will do some benchmarks and get back to
> you to understand it better.
>
> Thanks
> Govind
>
> > -----Original Message-----
> > From: Andrew 👽 Yourtchenko <ayour...@gmail.com<mailto:ayour...@gmail.com>>
> > Sent: Friday, March 27, 2020 7:52 AM
> > To: Govindarajan Mohandoss
> > <govindarajan.mohand...@arm.com<mailto:govindarajan.mohand...@arm.com>>
> > Cc: vpp-dev@lists.fd.io<mailto:vpp-dev@lists.fd.io>; nd
> > <n...@arm.com<mailto:n...@arm.com>>
> > Subject: Re: [vpp-dev] ACL question
> >
> > > On 27 Mar 2020, at 00:47, Govindarajan Mohandoss
> > <govindarajan.mohand...@arm.com<mailto:govindarajan.mohand...@arm.com>>
> > wrote:
> > >
> > >
> > >
> > > Hi Andrew,
> > >
> > > I just found out that ACL action differentiates SF or SL.
> > > Following
> > command enables SF and provides better performance.
> > >
> > > “acl_add_replace -1 ipv4 permit+reflect dst 192.82.1.1/32”
> > >
> > >
> > >
> > > Few more questions:
> > >
> > > =================
> > >
> > > Choosing between VPP Classifiers and ACL Plugin:
> > >
> > >
> > > https://lists.fd.io/g/vpp-dev/message/5716?p=,,,20,0,0,0::relevance,
> > > ,A
> > > CL,20,2,60,10641995
> > >
> > > You mentioned that VPP classifiers are faster than ACL plugin.
> > > For <L2, L3, L4> field based classification, which one provides
> > > better data
> > plane perf ?
> >
> >
> > It depends. If you wanna simultaneously match on all three, there is
> > currently no mechanism to generically do so.
> >
> > But then every time I looked at the use cases claiming to require
> > that, turned out it was a bad idea to represent the data this way -
> > because of combinatorial explosion. Even ACLs themselves suffer from
> > this issue - N sources times M destinations times K servces equal
> > N*M*K rules, which quickly skyrockets.
> >
> > > Does classifier support ranges ?
> >
> >
> > Classifier supports chained masked lookups. You might emulate ranges
> there.
> >
> > That said, I had seen ranges used only in a tiny percentage of the
> > cases. So they are a corner case imho.
> >
> >
> > > Which one is better if the rate of ACL rule add/del is high / low?
> >
> >
> > Classifier single table is your best bet probably. ACL plugin
> > deliberately does not have an API to add/del a single rule - you
> > always download the entire ACL.
> >
> > > Whether ACL rule priority is supported in both the schemes ?
> >
> >
> > First match for Acl and multi table classify case. Single table is
> > just a hash lookup because the entries don’t overlap by definition
> >
> > > Whether ACL Plugin SF mode will perform better than classifier ?
> >
> >
> > I did not benchmark them. It's somewhat different use cases.
> >
> > > Whether classifier also has SF mode ?
> >
> >
> > Nope.
> >
> > >
> > >
> > > ACL Plugin:
> > >
> > > SF mode – How much of extra memory is needed compared to SL mode ?
> >
> >
> > Depending on the number of active sessions... each session creates two
> > binash table entries, and consumes an entry in the session pool. The
> > default values in the code for the bihash memory usage have been
> > tested with half a million sessions - so you can extrapolate from
> > those with some ballpark (though bihash memory usage is not linear wrt
> > the entries, and also there is some extra memory churn due to bucket
> > reallocations when the size increases).
> >
> > —a
> >
> > >
> > >
> > > Thanks
> > >
> > > Govind
> > >
> > >
> > >
> > > From: vpp-dev@lists.fd.io<mailto:vpp-dev@lists.fd.io>
> > > <vpp-dev@lists.fd.io<mailto:vpp-dev@lists.fd.io>> On Behalf Of
> > > Govindarajan Mohandoss via Lists.Fd.Io
> > > Sent: Thursday, March 26, 2020 12:37 PM
> > > To: Andrew 👽 Yourtchenko <ayour...@gmail.com<mailto:ayour...@gmail.com>>
> > > Cc: vpp-dev@lists.fd.io<mailto:vpp-dev@lists.fd.io>
> > > Subject: Re: [vpp-dev] ACL question
> > >
> > >
> > >
> > > Hi Andrew,
> > >
> > > Thanks for the document.
> > >
> > > Can you please share the documents related to ACL plugin CLI
> > > config for
> > both stateful & stateless modes ?
> > >
> > >
> > >
> > > I tried the following commands for input ACL in VAT CLI. Not sure
> > whether this is SL / SF ?
> > >
> > > “
> > >
> > > vat# acl_add_replace -1 ipv4 permit dst 192.82.1.1/32
> > >
> > > vl_api_acl_add_replace_reply_t_handler:70: ACL index: 0
> > >
> > > vat# acl_interface_set_acl_list TenGigabitEthernet13/0/0 input 0
> > >
> > > vat# acl_interface_list_dump TenGigabitEthernet13/0/0
> > >
> > > vl_api_acl_interface_list_details_t_handler:115: sw_if_index: 3,
> > > count: 1, n_input: 1
> > >
> > > input 0
> > >
> > >
> > >
> > > vat# help acl_add_replace
> > >
> > > usage: acl_add_replace <acl-idx> [<ipv4|ipv6>]
> > <permit|permit+reflect|deny|action N> [src IP/plen] [dst IP/plen]
> > [sport X-Y] [dport X-Y] [proto P] [tcpflags FL MASK], ... , ...
> > >
> > > “
> > >
> > >
> > >
> > > Thanks
> > >
> > > Govind
> > >
> > >
> > >
> > > From: Andrew 👽 Yourtchenko <ayour...@gmail.com<mailto:ayour...@gmail.com>>
> > > Sent: Thursday, March 26, 2020 4:49 AM
> > > To: Govindarajan Mohandoss
> > > <govindarajan.mohand...@arm.com<mailto:govindarajan.mohand...@arm.com>>
> > > Cc: vpp-dev@lists.fd.io<mailto:vpp-dev@lists.fd.io>; Lijian Zhang
> > > <lijian.zh...@arm.com<mailto:lijian.zh...@arm.com>>;
> > > Jieqiang Wang <jieqiang.w...@arm.com<mailto:jieqiang.w...@arm.com>>; nd
> > > <n...@arm.com<mailto:n...@arm.com>>
> > > Subject: Re: [vpp-dev] ACL question
> > >
> > >
> > >
> > > As an acl plugin author I can say both stateful and stateless ACLs
> > > are used
> > for different consumers.
> > >
> > >
> > >
> > > Various matching implementations in vpp are used in different use
> cases...
> > and there is not a single silver bullet magic answer, because the
> > trade offs are different.
> > >
> > >
> > >
> > > https://nonsns.github.io/paper/rossi19ton.pdf
> > >
> > >
> > >
> > > Is a reasonable read on the subject - also because it relates to VPP
> > > and the
> > real project that we did a while ago.
> > >
> > >
> > >
> > > --a
> > >
> > >
> > >>
> > >> On 25 Mar 2020, at 17:26, Govindarajan Mohandoss
> > <govindarajan.mohand...@arm.com<mailto:govindarajan.mohand...@arm.com>>
> > wrote:
> > >>
> > >>
> > >>
> > >> Hello ACL Maintainer,
> > >>
> > >> We want to measure and optimize the ACL performance for ARM
> > servers. As per the foll. link, there are 4 different implementation
> > of ACLs in VPP.
> > >>
> > >> https://fd.io/docs/vpp/master/usecases/acls.html
> > >>
> > >> We would like to start with most commonly used ACL implementation
> > >> in
> > VPP which can cover L2, L3 and L4 fields. As per the link above and
> > CSIT reports (link below), it looks like ACL plugin is the right match.
> > >>
> > >> Can you please confirm ? ACL plugin has 2 variants – Stateful &
> Stateless.
> > Which is common and widely used in VPP ?
> > >>
> > >>
> > >> https://docs.fd.io/csit/master/report/detailed_test_results/vpp_per
> > >> fo
> > >> rmance_results/index.html
> > >>
> > >>
> > >>
> > >> Thanks
> > >>
> > >> Govind
> > >>
> > >> IMPORTANT NOTICE: The contents of this email and any attachments
> > >> are
> > confidential and may also be privileged. If you are not the intended
> > recipient, please notify the sender immediately and do not disclose
> > the contents to any other person, use it for any purpose, or store or
> > copy the information in any medium. Thank you.
-=-=-=-=-=-=-=-=-=-=-=-
Links: You receive all messages sent to this group.
View/Reply Online (#16185): https://lists.fd.io/g/vpp-dev/message/16185
Mute This Topic: https://lists.fd.io/mt/72544608/21656
Group Owner: vpp-dev+ow...@lists.fd.io
Unsubscribe: https://lists.fd.io/g/vpp-dev/unsub [arch...@mail-archive.com]
-=-=-=-=-=-=-=-=-=-=-=-
