racluster() memory control Re: New To Argus
Nick Diel
ndiel at engr.colostate.edu
Thu Feb 28 16:59:03 EST 2008
Carter,
I am going to start playing around with idle=timeout, if that parameter
is not specified is there a default value used or will all flows stay in
cache? Though this parameter looks very promising for my use.
Where we do most of our capturing we can see millions of port scans in a
12 hour trace, so that is an issue for us too when we do flow
filtering. I wonder if a separate timeout would be useful for flows
that just have a syn? Basically trying to purge out port scans faster.
Or maybe in a memory constraint model these flows are picked first to be
outputted.
I also think flushing out "complete" tcp flows is a good idea too.
Maybe a second timeout should be in place for these flows (it would be
shorter than the regular timeout and potentially 0), that way if you
wanted you could capture anomalies such as duplicate fin acks. This
second timer could also be used for flows that have a reset, since it is
very common to see additional packets after a reset (packets still on
the wire, reset gets lost, etc.).
Finally, I can see how a memory limit could be beneficial. While yes it
does create a problem where results are going to be influenced on size
of memory available, it will allow for processing that may not otherwise
be possible (or at least easily doable). When I was producing a list of
flows on port 25, I had to use very aggressive filters to handle memory
issues and I know I missed some flows anyways. We ended up with 20
million plus flows for our 12 hour capture. I would have been willing
to give a memory limit and had known that possibly not all flows would
have been combined properly. In my case at least I would expect most
flows outputted early due to memory constraints would have been port
scans and complete flows that haven't reached their idle timeout yet.
So again this would be a site specific option. Per your list:
1. filter input data
2. change the flow model definition to reduce the number of flows,
3. use the "idle=timeout" option on items in the racluster.conf file.
4. use memory limit for very large data sets with knowledge it could
affect the actual output.
Basically a memory limit is used when the others are not working. It
just allows for processing that may not have been easily possible.
Nick
Carter Bullard wrote:
> Hey Nick,
> I can put memory limits into racluster(), but then there is the
> possibililty
> that you get different results based on the available space. I'm not sure
> that is the right way to go, but who knows, it maybe a great option.
>
> The trick to keeping racluster memory use down is to:
> 1. filter input data
> 2. change the flow model definition to reduce the number of flows,
> 3. use the "idle=timeout" option on items in the racluster.conf file.
>
> This all needs to be customized for each site, so working with the
> racluster.conf file is the way to go, and running different .conf files
> against one sample test data file allows you fine tune the configuration.
>
> Getting darknet traffic out of the mix is important. For many sites
> "all the flows" are really scans, and should be ignored, 99.999%
> of the time. I track if something new responds to a scan, not that the
> scan exists, because there is always a scan, because many of the sites
> that I pay attention to have literally 100,000's of scans a day. As a
> result, we want to pay attention to the originator of the scan, the scan
> type, if the addresses involved are real, if its coming from inside or
> outside
> and if there was a "new" response. Sloughing scan traffic to tools
> that do scan analysis, and tracking the other flows makes this
> a doable thing, and programs like radark()/rafilteraddr() help here
> (but they are just examples).
>
> For traffic you really want to track, modifying the flow model allows
> us to reduce the number of flow caches, say, by ignoring the source
> port for flows going to well known servers. Lines like this:
>
> filter="dst host 1.2.3.4 and src pkts eq 1 and dst pkts eq 1 and
> dst port 53" model="saddr/16 daddr proto dport"
>
> will reduce the number of in memory caches for this DNS server to
> just the number of class B networks hitting the server. The filter
> only applies to successful connections, and the resulting aggregation
> record will contain stats like the number of connections and avgdur.
> (I'm assuming here that you will run racluster() against a file).
>
> You can have literally 1000's of these filters to generate the data
> you really want.
>
> When a flow hits its idle time, racluster will write out the record (if
> needed) and deallocate the memory used for that flow. So having
> aggressive idle times is very helpful.
>
> But now that I'm thinking about it, we don't really have a way of
> flushing records based on the flow state for flows that are being merged
> together. What I mean by that, is that if a TCP connection has 12 status
> records, and we finally get the last one that "closes" the connection,
> there isn't a way, currently, for us to check to see if that flow should
> be "flushed".
>
> Possibly we should take the resulting merged record, and run it back
> through the filters to see if there is something we need to do with it.
>
> Well, anyway, keep sending email if any of this is useful.
>
> Carter
>
>
>
> On Feb 28, 2008, at 1:25 PM, Nick Diel wrote:
>
>> Carter,
>>
>> Thanks for all of your input. Also thanks for the updated Argus.
>>
>> After reading what you said, I can understand why Argus was designed
>> the way it was. I was just initially evaluating Argus with some very
>> simple and discrete examples. Looking at some of the source code
>> also helped me wrap my head around Argus.
>>
>> On to the memory issue. The system I am using has 2GB in it and
>> racluster wants to use all of it. When 1.7GB< starts to get used by
>> racluster heavy swapping occurs and racluster's cpu usage drops below
>> 25%. So this was why I was thinking out loud about potentially
>> giving racluster a memory limit from the command line. This way the
>> system could avoid the heavy swapping and just have racluster write
>> out the oldest records before moving on.
>>
>> Again thanks for putting up with me as I start to understand Argus.
>>
>> Nick
>>
>> Carter Bullard wrote:
>>> Hey Nick,
>>> The problem with packet capture, is primarily the disk performance.
>>> Argus can go as fast as you can collect packets, assuming that
>>> you're using Endace cards, and although argus does great in the
>>> presence of packet loss, it generates its best results when it gets
>>> everything.
>>>
>>> The best architecture is to run argus on the packet capture box,
>>> and to blow argus records to another machine that does the disk
>>> operations to store records. This division of labor works best for
>>> the 10Gbps capture facilities.
>>>
>>> We sort input file names in the ra* programs, so doing this for
>>> argus is a cut, copy, paste job. No problem, I'll put it in this week.
>>>
>>> Argus can read from stdin.
>>>
>>> There are many incantations that work to decrease the memory
>>> demands of an argus client. Just really need to know what it is
>>> that you want to do.
>>>
>>> OK, to your question.
>>>>
>>>> Now let me ask about what I have been working on (merging flows
>>>> across argus data files). First, if I was capturing with Argus
>>>> (not reading pcap files, capturing off the wire: argus | raspilt)
>>>> wouldn't I run into the same problem of having flows broken up
>>>> across different argus files?
>>>>
>>>> If racluster is merging records as it finds them (not reading all
>>>> records into memory first), it seems it might be nice to specify a
>>>> memory limit for racluster at command line. Then as racluster
>>>> approaches the memory limit it could remove the oldest records from
>>>> memory and print them to the output.
>>>
>>> Multiple argus status records spanning files. Well, yes that is the
>>> actual design
>>> goal. When you think about most types of
>>> operations/security/performance
>>> analysis, you want to see flow data scoped over some time boundary.
>>> Regardless
>>> of what that boundary is, whether its the last millisecond, second
>>> or minute or hour,
>>> you will have flows that span that boundary. There are a lot of
>>> flows that are
>>> persistent, so you can't have a file big enough to hold complete
>>> flows, ....,
>>> really.
>>>
>>> But you don't seem to be too interested in really granular data, so
>>> you should
>>> modify the ARGUS_FAR_STATUS_INTERVAL value to be something larger
>>> than your file duration. That way argus generates only one record
>>> per flow
>>> per file. You use ra() to split files that are complete from those
>>> that may
>>> continue into the next file, using the "-E" option and then after
>>> you're done
>>> with all the files you have, then run racluster() against these
>>> continuation files.
>>>
>>> for i in *pcap; do argus -S 5m -r $i -w $i.argus; done
>>> for i in *argus; do ra -r $i -E $i.cont -w argus.out - tcp and
>>> ((syn or synack) and (fin or finack or reset)); done
>>> racluster -r *.cont -w argus.out
>>>
>>> They won't be sorted, but thats easy to do with an additional step:
>>> rasplit -M nomodify -r argus.out -M time 5m -w
>>> data/argus.%Y.%m.%d.%H.%M.%S
>>> rm argus.out
>>> rasort -R data -M replace
>>> ra -R data -w argus.out
>>> rm -rf data
>>>
>>> Or at least something like that should work. The "-M nomodify" is
>>> critical, as rasplit()
>>> with break records up into time boundaries if you don't specify this
>>> option, which
>>> puts you back in trouble, if you're really trying to keep the flows
>>> together.
>>>
>>> Argus clients aren't suppose to consume more than, what 1GB of
>>> memory, so there
>>> are limits in the code. Do you have a smaller machine than that?
>>>
>>>
>>> Carter
>>>
>>>
>>> On Feb 25, 2008, at 2:01 PM, Nick Diel wrote:
>>>
>>>> Carter,
>>>>
>>>> First of all thanks for your detailed response and updated
>>>> clients. And I am glad you like twists.
>>>>
>>>> Let me tell you a little bit more about the research setup. The
>>>> research project I am part of (made up of several universities in
>>>> the US) has several collection boxes in different large commercial
>>>> environments. The boxes were customized specifically for high
>>>> speed packet capturing (RAID, Endace capture card, etc.). We will
>>>> run a 12 hour capture and then analyze the capture for some time.
>>>> Sometimes up to several months. So I do have time to correctly
>>>> create my argus output files and do any other processing I need to do.
>>>>
>>>> Some of the researchers focus on packet based research, where as
>>>> other parts of the group focus more on flow based analysis. So
>>>> Argus looks like a great match for us. Immediately after the
>>>> capture, we can create Argus flow records and do our flow analysis
>>>> with Argus clients.
>>>>
>>>> So for my first question, is Argus capable of capturing at high
>>>> line speeds (at least 1Gbit) where doing a packet capture using
>>>> libpcap and a standard NIC may fail (libpcap dropping packets)? Or
>>>> since Argus is flow based it doesn't care if it misses packets?
>>>> Some of the anomalies we research require us to account for almost
>>>> every packet in the anomaly, so say dropping every 100th or even
>>>> every 1000th packet could hamper us. The reason I ask I about
>>>> Argus high speed captures, is if it is very capable at high speeds,
>>>> it would allow us to deploy more collection boxes (these boxes
>>>> would then primarily be used by the flow based researchers). We
>>>> wouldn't have to buy an expensive capture card for each collection box.
>>>>
>>>> As for reading multiple files into Argus, one easy way to
>>>> accomplish this would have Argus be able to read pcap files from
>>>> stdin. Then one can use a utility such as mergecap or tcpslice to
>>>> feed Argus a list of out of order files: mergecap -r
>>>> /packets/*.pcap -w - | argus -r - ....
>>>>
>>>> My files are named so chronological order equals lexical order so
>>>> argus -r * would work in my case (this helps us with a number of
>>>> utilities we use). I do understand actually implementing this in
>>>> Argus would require probably a number of things such as dieing when
>>>> files are out of order and then telling the user what order argus
>>>> was reading the files. Though doing this would be quite faster
>>>> then having tcpslice or mergecap feed Argus the pcap files.
>>>>
>>>> Now let me ask about what I have been working on (merging flows
>>>> across argus data files). First, if I was capturing with Argus
>>>> (not reading pcap files, capturing off the wire: argus | raspilt)
>>>> wouldn't I run into the same problem of having flows broken up
>>>> across different argus files?
>>>>
>>>> If racluster is merging records as it finds them (not reading all
>>>> records into memory first), it seems it might be nice to specify a
>>>> memory limit for racluster at command line. Then as racluster
>>>> approaches the memory limit it could remove the oldest records from
>>>> memory and print them to the output.
>>>>
>>>> I was able to use your suggestion successfully to merge most of my
>>>> flows together. Though I needed to make a few modifications to the
>>>> filter. I moved parenthesis, "tcp and ((syn or synack) and (*(*fin
>>>> or finack) or reset*)*)" vs. "tcp and (*(*(syn or synack) and (fin
>>>> or finack)*)* or reset)." And I added "not con" to filter out the
>>>> many, many packet scans, though this also does not merge syn-synack
>>>> flows which exist at the end of the argus output files. This
>>>> filter still caused most of the memory to be used, but not a whole
>>>> lot of time was spent in the upper range where swapping was slowing
>>>> the system to a crawl. Without "not con" I would reach the upper
>>>> limits of memory usage quite fast and go into a crawl with the
>>>> swapping.
>>>>
>>>> Thanks again for all your help,
>>>> Nick
>>>>
>>>>
>>>> Carter Bullard wrote:
>>>>> Hey Nick,
>>>>> The argus project from the very beginning has been trying
>>>>> to get people away from capturing packets, and instead
>>>>> capturing comprehensive flow records that account for every
>>>>> packet on the wire. This is because capturing packets at modern
>>>>> speeds seems impractical, and there are a lot of problems that can
>>>>> be worked out without all that data.
>>>>>
>>>>> So to use argus in the way you want to use argus is a bit of a
>>>>> twist on the model. But I like twists ;o)print
>>>>>
>>>>> >>> To start out with something simple I want to be able to count
>>>>> the number of flows over TCP port 25.
>>>>>
>>>>> The easiest way to do that right now is to do something like this
>>>>> in bash:
>>>>>
>>>>> % for i in pcap*; do argus -r $i -w - - tcp and port 25 | \
>>>>> rasplit -M time 5m -w -
>>>>> argus.data/%Y/%m/%d/argus.%Y.%m.%d.%H.%M.%S ; \
>>>>> done
>>>>>
>>>>> That will put the tcp:25 "micro flow" argus records into a
>>>>> manageable
>>>>> set of files. Now the files themselves need to be processed to
>>>>> get the flows merged together:
>>>>>
>>>>> % racluster -M replace -R argus.data
>>>>>
>>>>> So now you'll get the data needed to ask questions, split into 5m
>>>>> bins,
>>>>> so to speak. Changing the "5m" to "1h", "4h", or "1d", may generate
>>>>> file structures that you can work with, but eventually you will
>>>>> hit a memory
>>>>> wall. Without doing something clever.
>>>>>
>>>>> Now that you have these intermediate files, in order to merge the
>>>>> tcp flows that span multiple files, you will need to give racluster()
>>>>> a different aggregation strategy than the default. Try a
>>>>> racluster.conf file that contains these lines against the argus files
>>>>> you have.
>>>>>
>>>>> ------- start racluster.conf ---------
>>>>>
>>>>> filter="tcp and ((syn or synack) and ((fin or finack) or reset))"
>>>>> status=-1 idle=0
>>>>> filter="" model="saddr daddr proto sport dport"
>>>>>
>>>>> ------- end racluster.conf --------
>>>>>
>>>>> What this will do is:
>>>>> 1. any tcp connection that is complete, where we saw the
>>>>> beginning and the
>>>>> end, just pass it through, don't track anything.
>>>>> 2. any partial tcp connection, track and merge records that match.
>>>>>
>>>>> So it only allocates memory for flows that are 'continuation'
>>>>> records.
>>>>> The output is unsorted, so you will need to run rasort() if you want
>>>>> to do any time oriented operations on the output.
>>>>>
>>>>> In testing this, I found a problem with parsing "-1" from the status
>>>>> field in some weird conditions, so I fixed it. Grab the newest
>>>>> clients from the dev directory if you want to try this method.
>>>>>
>>>>> ftp://qosient.com/dev/argus-3.0/argus-clients-3.0.0.rc.69.tar.gz
>>>>>
>>>>> Give that a try, and send email to the list with any kind of result
>>>>> yiou get.
>>>>>
>>>>> With so many pcap files, we probably need to make some other
>>>>> changes.
>>>>>
>>>>> The easiest way for you to do what you eventually want do,
>>>>> would be for you to say something like this:
>>>>> argus -r * -w - | rawhatever
>>>>>
>>>>> This current won't work, and there is a reason, but maybe we
>>>>> can change it. Argus currently can read multiple input files, but
>>>>> you
>>>>> need to specify each file using a "-r filename -r filename " like
>>>>> command
>>>>> line list. With 1000's of files, that is somewhat impractical.
>>>>> It is this
>>>>> way on purpose, because argus really does need to see packets in
>>>>> time order.
>>>>>
>>>>> If you try to do something like this:
>>>>>
>>>>> argus -r * -w - | rasplit -M time 5m -w
>>>>> argus.out.%Y.%m.%d.%H.%M.%S
>>>>>
>>>>> which is designed generate argus record files that represent packet
>>>>> behavior with hard cutoffs every 5 minutes, on the hour; if the
>>>>> packet files are not read in time order, you get really weird
>>>>> results. It's as if the realtime argus was jumping into the
>>>>> future and
>>>>> then into the past and then back to the future again.
>>>>>
>>>>> Now, if you name your pcap files so they can be sorted, I can
>>>>> make it so "argus -r *" can work. How do you name your pcap files?
>>>>>
>>>>>
>>>>> Because argus has the same timestamps as the packets in your
>>>>> pcap files, the timestamps can be used as an "external key" if
>>>>> you will. If you build a database that has tuples (entries) like:
>>>>>
>>>>> "pcap_filename start_time end_time"
>>>>>
>>>>> then by looking at a single argus record, which has a start time
>>>>> and an end time, you can find the pcap files that contain its
>>>>> packets.
>>>>> And with something like perl and tcpdump or wireshark, you can
>>>>> feed a simple shell to look in those pcap files looking for packets
>>>>> with this type of filter:
>>>>>
>>>>> ( ether host $smac and $dmac) and (host $saddr and $daddr) and
>>>>> ports \
>>>>> ($sport and $dport)
>>>>>
>>>>> and you get all the packets that are referenced in the record.
>>>>>
>>>>>
>>>>> Carter
>>>>>
>>>>>
>>>>>
>>>>>
>>>>> On Feb 21, 2008, at 4:49 PM, Nick Diel wrote:
>>>>>
>>>>>> I am new to Argus, but have found it has great potential for the
>>>>>> research project I work on. We collect pcap files from several
>>>>>> high traffic networks (20k-100k packets/second). We collect for
>>>>>> approximately 12 hours and have ~1000 pcap files that are roughly
>>>>>> 500MB each.
>>>>>> I am wanting to do a number of different flow analysis and think
>>>>>> Argus might be perfect for me. I am having a hard time grasping
>>>>>> some of the fundamentals of Argus, but I think once I get some of
>>>>>> the basics I will be able to really start to use Argus.
>>>>>>
>>>>>> To start out with something simple I want to be able to count the
>>>>>> number of flows over TCP port 25. I know I need to use RACluster
>>>>>> to merge the Argus output (I have one argus file for each pcap
>>>>>> file I have), that way I can combine identical flow records into
>>>>>> one. I can do this fine on one argus output file, but I know
>>>>>> many flows span the numerous files I have. I also know I can't
>>>>>> load all the files at once into RACluster as it fills all
>>>>>> available memory. So my question is how can I accomplish this
>>>>>> while making sure I capture most flows that span multiple files.
>>>>>>
>>>>>> Once I understand this, I hope to be able to do things like
>>>>>> create a list of flow sizes (in bytes) for port 25. Basically I
>>>>>> will be asking a lot of questions involving all flows that match
>>>>>> a certain filter and I am not sure how to accommodate for flows
>>>>>> spanning multiple files.
>>>>>>
>>>>>> A separate question. I don't think Argus has this ability, but I
>>>>>> wanted to know if the community already had a utility for this.
>>>>>> I am looking into creating a DB of some sort that would match
>>>>>> Argus's flow IDs to pcap file name(s) and packet numbers. This
>>>>>> way one could extract the packets for a flow that needed further
>>>>>> investigation.
>>>>>>
>>>>>> And finally, thanks for the great tool. It does a number of
>>>>>> things I have been doing manually for a while.
>>>>>>
>>>>>> Thanks,
>>>>>> Nick
>>>>>>
>>>>>>
>>>>>
>>>>
>>>
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://pairlist1.pair.net/pipermail/argus/attachments/20080228/5935ee0a/attachment.html>
More information about the argus
mailing list