From: Pablo Neira Ayuso <pablo@netfilter.org>
To: Phil Sutter <phil@nwl.cc>, netfilter-devel@vger.kernel.org
Subject: Re: [iptables PATCH v3 0/7] Improve xtables-restore performance
Date: Wed, 6 Nov 2019 10:24:52 +0100 [thread overview]
Message-ID: <20191106092452.2witubxzularwbn2@salvia> (raw)
In-Reply-To: <20191031171947.GF8531@orbyte.nwl.cc>
Hi Phil,
On Thu, Oct 31, 2019 at 06:19:47PM +0100, Phil Sutter wrote:
> On Thu, Oct 31, 2019 at 04:02:34PM +0100, Pablo Neira Ayuso wrote:
> > On Thu, Oct 24, 2019 at 06:37:05PM +0200, Phil Sutter wrote:
> > > This series speeds up xtables-restore calls with --noflush (typically
> > > used to batch a few commands for faster execution) by preliminary input
> > > inspection.
> > >
> > > Before, setting --noflush flag would inevitably lead to full cache
> > > population. With this series in place, if input can be fully buffered
> > > and no commands requiring full cache is contained, no initial cache
> > > population happens and each rule parsed will cause fetching of cache
> > > bits as required.
> > >
> > > The input buffer size is arbitrarily chosen to be 64KB.
> > >
> > > Patches one and two prepare code for patch three which moves the loop
> > > content parsing each line of input into a separate function. The
> > > reduction of code indenting is used by patch four which deals with
> > > needless line breaks.
> >
> > For patches from 1 to 4 in this batch:
> >
> > Acked-by: Pablo Neira Ayuso <pablo@netfilter.org>
> >
> > > Patch five deals with another requirement of input buffering, namely
> > > stripping newline characters from each line. This is not a problem by
> > > itself, but add_param_to_argv() replaces them by nul-chars and so
> > > strings stop being consistently terminated (some by a single, some by
> > > two nul-chars).
> > >
> > > Patch six then finally adds the buffering and caching decision code.
> > >
> > > Patch seven is pretty unrelated but tests a specific behaviour of
> > > *tables-restore I wasn't sure of at first.
> >
> > Do you have any number?
>
> Yes, I wrote a small benchmark based on some Kubernetes use-case. It
> measures loading of dumps like:
>
> | *nat
> | :KUBE-SVC-23 - [0:0]
> | :KUBE-SEP-23 - [0:0]
> | -A KUBE-HOOK ! -s 10.128.0.0/14 -d 172.30.108.136/32 -p tcp -m comment --comment \"openshift-controller-manager/controller-manager:https cluster IP\" -m tcp --dport 443 -j KUBE-MARK-MASQ
> | -A KUBE-HOOK -d 172.30.108.136/32 -p tcp -m comment --comment \"openshift-controller-manager/controller-manager:https cluster IP\" -m tcp --dport 443 -j KUBE-SVC-23
> | -A KUBE-SVC-23 -j KUBE-SEP-23
> | -A KUBE-SEP-23 -s 10.128.0.38/32 -j KUBE-MARK-MASQ
> | -A KUBE-SEP-23 -p tcp -m tcp -j DNAT --to-destination 10.128.0.38:8443
> | COMMIT
>
> Into a ruleset with increasing size (created by repeating the snippet above):
>
> size (*100) | legacy | nft-pre | nft-post
> ---------------------------------------------------------
> 1 .0040366426 .0079313714 .0025598650
> 10 .0146918664 .0459193868 .0025134858
> 25 .0361553334 .1195503778 .0024202904
> 50 .0699177362 .2547542626 .0024351612
> 75 .1062593206 .4078182120 .0024362044
> 100 .1614045514 .5636617378 .0024195190
Thanks, this is nice.
I can see this function:
static bool cmd_needs_full_cache(char *cmd)
the pre-parsing of the input to calculate the cache, which is good.
One thing: why do you need the conversion from \n to \0. The idea is
to read once from the file and keep it in a buffer, then pass it to
the original parsing function after this pre-parsing to calculate the
cache.
Please, add this to the remaining patches of this series.
Acked-by: Pablo Neira Ayuso <pablo@netfilter.org>
Thanks.
next prev parent reply other threads:[~2019-11-06 9:24 UTC|newest]
Thread overview: 12+ messages / expand[flat|nested] mbox.gz Atom feed top
2019-10-24 16:37 [iptables PATCH v3 0/7] Improve xtables-restore performance Phil Sutter
2019-10-24 16:37 ` [iptables PATCH v3 1/7] xtables-restore: Integrate restore callbacks into struct nft_xt_restore_parse Phil Sutter
2019-10-24 16:37 ` [iptables PATCH v3 2/7] xtables-restore: Introduce struct nft_xt_restore_state Phil Sutter
2019-10-24 16:37 ` [iptables PATCH v3 3/7] xtables-restore: Introduce line parsing function Phil Sutter
2019-10-24 16:37 ` [iptables PATCH v3 4/7] xtables-restore: Remove some pointless linebreaks Phil Sutter
2019-10-24 16:37 ` [iptables PATCH v3 5/7] xtables-restore: Allow lines without trailing newline character Phil Sutter
2019-10-24 16:37 ` [iptables PATCH v3 6/7] xtables-restore: Improve performance of --noflush operation Phil Sutter
2019-10-24 16:37 ` [iptables PATCH v3 7/7] tests: shell: Add ipt-restore/0007-flush-noflush_0 Phil Sutter
2019-10-31 15:02 ` [iptables PATCH v3 0/7] Improve xtables-restore performance Pablo Neira Ayuso
2019-10-31 17:19 ` Phil Sutter
2019-11-06 9:24 ` Pablo Neira Ayuso [this message]
2019-11-06 12:31 ` Phil Sutter
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=20191106092452.2witubxzularwbn2@salvia \
--to=pablo@netfilter.org \
--cc=netfilter-devel@vger.kernel.org \
--cc=phil@nwl.cc \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).