fuzzy subnet aggregation
Does anyone have or seen any such tool? I have a script that seems to work, but its terribly slow. Currently I can produce aggregated subnets that can be mising up to a specified number of individual addresses. Which can be fed back in for multiple passes. Doing RTBH on individual /32 does not scale well, if you are eyeing collaboration with external lists. I have found likely sources that could produce another 100k prefixes easily. Joe
Is this what you are trying to accomplish? $ python Python 2.7.15rc1 (default, Nov 12 2018, 14:31:15) [GCC 7.3.0] on linux2 Type "help", "copyright", "credits" or "license" for more information.
import netaddr SomeList=netaddr.IPSet() SomeList.add('203.0.113.0/25') SomeList.add('203.0.113.128/25') for x in list(SomeList.iter_cidrs()): ... print x ... 203.0.113.0/24
DifferentList=netaddr.IPSet() DifferentList.add('0.0.0.0/0') DifferentList.remove('203.0.113.1') for x in list(DifferentList.iter_cidrs()): ... print x ... 0.0.0.0/1 128.0.0.0/2 192.0.0.0/5 200.0.0.0/7 202.0.0.0/8 203.0.0.0/18 203.0.64.0/19 203.0.96.0/20 203.0.112.0/24 203.0.113.0/32 203.0.113.2/31 203.0.113.4/30 203.0.113.8/29 203.0.113.16/28 203.0.113.32/27 203.0.113.64/26 203.0.113.128/25 203.0.114.0/23 203.0.116.0/22 203.0.120.0/21 203.0.128.0/17 203.1.0.0/16 203.2.0.0/15 203.4.0.0/14 203.8.0.0/13 203.16.0.0/12 203.32.0.0/11 203.64.0.0/10 203.128.0.0/9 204.0.0.0/6 208.0.0.0/4 224.0.0.0/3
On Sun, Oct 27, 2019 at 1:10 PM Joe Maimon <jmaimon@jmaimon.com> wrote:
Does anyone have or seen any such tool? I have a script that seems to work, but its terribly slow.
Currently I can produce aggregated subnets that can be mising up to a specified number of individual addresses. Which can be fed back in for multiple passes.
Doing RTBH on individual /32 does not scale well, if you are eyeing collaboration with external lists. I have found likely sources that could produce another 100k prefixes easily.
Joe
Not quite. 203.0.113.1 203.0.113.3 203.0.113.5 203.0.112.6 203.0.112.7 Will aggregate to 203.0.113.0/29 if you dont mind the missing 3 addresses in the unaggregated list. Hence, fuzzy aggregation. Joe
Is this what you are trying to accomplish?
$ python Python 2.7.15rc1 (default, Nov 12 2018, 14:31:15) [GCC 7.3.0] on linux2 Type "help", "copyright", "credits" or "license" for more information.
import netaddr SomeList=netaddr.IPSet() SomeList.add('203.0.113.0/25') SomeList.add('203.0.113.128/25') for x in list(SomeList.iter_cidrs()): ... print x ... 203.0.113.0/24
DifferentList=netaddr.IPSet() DifferentList.add('0.0.0.0/0') DifferentList.remove('203.0.113.1') for x in list(DifferentList.iter_cidrs()): ... print x ... 0.0.0.0/1 128.0.0.0/2 192.0.0.0/5 200.0.0.0/7 202.0.0.0/8 203.0.0.0/18 203.0.64.0/19 203.0.96.0/20 203.0.112.0/24 203.0.113.0/32 203.0.113.2/31 203.0.113.4/30 203.0.113.8/29 203.0.113.16/28 203.0.113.32/27 203.0.113.64/26 203.0.113.128/25 203.0.114.0/23 203.0.116.0/22 203.0.120.0/21 203.0.128.0/17 203.1.0.0/16 203.2.0.0/15 203.4.0.0/14 203.8.0.0/13 203.16.0.0/12 203.32.0.0/11 203.64.0.0/10 203.128.0.0/9 204.0.0.0/6 208.0.0.0/4 224.0.0.0/3
On Sun, Oct 27, 2019 at 1:10 PM Joe Maimon <jmaimon@jmaimon.com> wrote:
Does anyone have or seen any such tool? I have a script that seems to work, but its terribly slow.
Currently I can produce aggregated subnets that can be mising up to a specified number of individual addresses. Which can be fed back in for multiple passes.
Doing RTBH on individual /32 does not scale well, if you are eyeing collaboration with external lists. I have found likely sources that could produce another 100k prefixes easily.
Joe
On 27. Oct 2019, at 20:36, Joe Maimon <jmaimon@jmaimon.com> wrote:
Not quite.
203.0.113.1 203.0.113.3 203.0.113.5 203.0.112.6 203.0.112.7
Will aggregate to 203.0.113.0/29 if you dont mind the missing 3 addresses in the unaggregated list.
Hence, fuzzy aggregation.
Could you describe the problem again? I’m interested, but I’m not sure that I quite understand what you want to do :-) were the last two addresses supposed to have 112 in the third octet? Nick
On Sun, Oct 27, 2019 at 3:09 PM Joe Maimon <jmaimon@jmaimon.com> wrote:
Does anyone have or seen any such tool? I have a script that seems to work, but its terribly slow.
Currently I can produce aggregated subnets that can be mising up to a specified number of individual addresses. Which can be fed back in for multiple passes.
your aim is to get to maximum aggregation .. with some overage, like 90% of a /24 ? so missing like 25 addresses in a whole /24.. (for instance)
Doing RTBH on individual /32 does not scale well, if you are eyeing collaboration with external lists. I have found likely sources that could produce another 100k prefixes easily.
Joe
On Sun, Oct 27, 2019 at 3:09 PM Joe Maimon <jmaimon@jmaimon.com> wrote:
your aim is to get to maximum aggregation .. with some overage, like 90% of a /24 ? so missing like 25 addresses in a whole /24.. (for instance)
I would be happy to get /29's missing 3 /28's missing 5, etc... This is not punitive, its about scale. Joe
Are you trying to reduce the number of ACL rules that include a known set of addresses but also minimize covered addresses that are not part of the mandatory set? Tony
On Oct 27, 2019, at 12:29, Joe Maimon <jmaimon@jmaimon.com> wrote:
On Sun, Oct 27, 2019 at 3:09 PM Joe Maimon <jmaimon@jmaimon.com> wrote:
your aim is to get to maximum aggregation .. with some overage, like 90% of a /24 ? so missing like 25 addresses in a whole /24.. (for instance)
I would be happy to get /29's missing 3 /28's missing 5, etc...
This is not punitive, its about scale.
Joe
So I went back to the drawing board, and I think I have something that seems to work much better. - convert input prefixes to single ip expressed as integer - sort -n | uniq - into a temporary list file begin read sequentially until maxhosts (or minhosts) or next subnet If matched enough single addresses, output subnet (and missing hosts without early loop termination) delete all subnet addresses read loop Total process time on a vm on old hardware, less than 2m for a 5500 line input. Now to verify results, positive and negative.... Results are still raw, but anyone who wishes is welcome to it. Joe Joe Maimon wrote:
Does anyone have or seen any such tool? I have a script that seems to work, but its terribly slow.
Currently I can produce aggregated subnets that can be mising up to a specified number of individual addresses. Which can be fed back in for multiple passes.
Doing RTBH on individual /32 does not scale well, if you are eyeing collaboration with external lists. I have found likely sources that could produce another 100k prefixes easily.
Joe
You could modify a radix tree to include a consolidation function and resulting confidence. Then walk the nodes of the tree, check to see if the subtree meets the requirements for consolidation, if so, prune and record the confidence. You would need to re-run the consolidation from the original data every time an individual IP was added/removed from the list as the consolidation function is lossy. Alternatively, you could do consolidation on the fly losslessly if you had a custom tree walk algorithm. That's probably the way I would do it. I'm not a programmer so I assume there are better ways out there. Your processing time for 5k IPs should be measured in seconds (ie: less than one) rather than minutes on any modern core. Based on your pseudocode (sort -n | uniq) I get the impression that you're using BASH which isn't ideal for performing this sort of operation at high speed. On the flip side, I think an extra 100k routes isn't that much unless you're suffering from hardware routing table limitations. In my world the cost of a false positive match would far outweigh the cost of upgrading hardware. YMMV. Do you have a git repo? On Sun, Oct 27, 2019 at 9:58 PM Joe Maimon <jmaimon@jmaimon.com> wrote:
So I went back to the drawing board, and I think I have something that seems to work much better.
- convert input prefixes to single ip expressed as integer - sort -n | uniq - into a temporary list file
begin
read sequentially until maxhosts (or minhosts) or next subnet
If matched enough single addresses, output subnet (and missing hosts without early loop termination)
delete all subnet addresses read
loop
Total process time on a vm on old hardware, less than 2m for a 5500 line input. Now to verify results, positive and negative....
Results are still raw, but anyone who wishes is welcome to it.
Joe
Joe Maimon wrote:
Does anyone have or seen any such tool? I have a script that seems to work, but its terribly slow.
Currently I can produce aggregated subnets that can be mising up to a specified number of individual addresses. Which can be fed back in for multiple passes.
Doing RTBH on individual /32 does not scale well, if you are eyeing collaboration with external lists. I have found likely sources that could produce another 100k prefixes easily.
Joe
At this time I am just trying to get an idea if the whole exercise is worth it. Whether the processing time is feasible for 5k, 50k, 100k, 200k. Whether the results reduce the count measurably at acceptable collateral levels. Because rtbh scaling to 100k is one thing. And from there it could go higher. And it works best if it can be spread to as many edge devices as possible, including ones with limited tcam, such as customer edge l3 switches. And to customers/friends who have redundant connections with other providers, including broadband, possibly with tiny routers, even ddwrt. 50k routes here and there and soon you are talking about real table bloat. I try to start every project as a script if at all possible. If that works even somewhat real code is promising. Joe Mark Leonard wrote:
You could modify a radix tree to include a consolidation function and resulting confidence. Then walk the nodes of the tree, check to see if the subtree meets the requirements for consolidation, if so, prune and record the confidence. You would need to re-run the consolidation from the original data every time an individual IP was added/removed from the list as the consolidation function is lossy.
Alternatively, you could do consolidation on the fly losslessly if you had a custom tree walk algorithm. That's probably the way I would do it. I'm not a programmer so I assume there are better ways out there.
Your processing time for 5k IPs should be measured in seconds (ie: less than one) rather than minutes on any modern core. Based on your pseudocode (sort -n | uniq) I get the impression that you're using BASH which isn't ideal for performing this sort of operation at high speed.
On the flip side, I think an extra 100k routes isn't that much unless you're suffering from hardware routing table limitations. In my world the cost of a false positive match would far outweigh the cost of upgrading hardware. YMMV.
Do you have a git repo?
On Sun, Oct 27, 2019 at 9:58 PM Joe Maimon <jmaimon@jmaimon.com <mailto:jmaimon@jmaimon.com>> wrote:
So I went back to the drawing board, and I think I have something that seems to work much better.
- convert input prefixes to single ip expressed as integer - sort -n | uniq - into a temporary list file
begin
read sequentially until maxhosts (or minhosts) or next subnet
If matched enough single addresses, output subnet (and missing hosts without early loop termination)
delete all subnet addresses read
loop
Total process time on a vm on old hardware, less than 2m for a 5500 line input. Now to verify results, positive and negative....
Results are still raw, but anyone who wishes is welcome to it.
Joe
Joe Maimon wrote: > Does anyone have or seen any such tool? I have a script that seems to > work, but its terribly slow. > > Currently I can produce aggregated subnets that can be mising up to a > specified number of individual addresses. Which can be fed back in for > multiple passes. > > Doing RTBH on individual /32 does not scale well, if you are eyeing > collaboration with external lists. I have found likely sources that could > produce another 100k prefixes easily. > > Joe > >
Mark Leonard wrote : Your processing time for 5k IPs should be measured in seconds (ie: less than one) rather than minutes on any modern core.
I agree. I am surprised by the minutes thing as well.
Based on your pseudocode (sort -n | uniq) I get the impression that you're using BASH which isn't ideal for performing this sort of operation at high speed.
Even with bash it does not take minutes. I used a bash sort in the first version of the CBBC and it was quite fast. Was not very elegant, but worked fine. Fetch all sources into the same file, then bash sort removing duplicates. My current limitation now is the injection into ExaBGP and because I still display the prefixes in the console. Kinda, its in debug mode.
On the flip side, I think an extra 100k routes isn't that much unless you're suffering from hardware routing table limitations. In my world the cost of a false positive match would far outweigh the cost of upgrading hardware. YMMV.
The problem here is that there still are plenty of routers out there that have a 1 million route limit and that the growth of the routing table is predictable enough that the year to upgrade is 2023. By adding 100K prefixes, you advance the upgrade year into 2021. Does not look good to request the capex 2 years earlier than previously predicted. I am curious to see how many prefixes aggregation saves. Michel. TSI Disclaimer: This message and any files or text attached to it are intended only for the recipients named above and contain information that may be confidential or privileged. If you are not the intended recipient, you must not forward, copy, use or otherwise disclose this communication or the information contained herein. In the event you have received this message in error, please notify the sender immediately by replying to this message, and then delete all copies of it from your system. Thank you!...
Good news, bad news. With an inefficient bash script on an inefficient platform, 120k processes in less than 15minutes. Thus far, the best I have is less than 10% reduction with barely acceptable aggressiveness. The distribution is too varied or the level of aggressiveness has to be beyond any subtleties. Putting this one back on the shelf. Joe Michel Py wrote:
I am curious to see how many prefixes aggregation saves.
Michel.
participants (8)
-
Antonio Querubin -
Christopher Morrow -
Grant Taylor -
Joe Maimon -
Mark Leonard -
Masataka Ohta -
Michel Py -
Nick Morrison