Results 1 to 17 of 17
  1. #1
    Join Date
    Dec 2009
    Posts
    88

    95th percentile methods?

    Hello all,

    Currently looking into different methods of 95th percentile, just wondering what 95th percentile methods are you or other providers using?

    What is the sampling rate?
    Do you take the highest in or out per every sample then take the 95th of that data?
    Do you use in and out separately first, then take the higher of these two?
    Do you sum the total bandwidth (in and out) per sample and then take the 95th of that data?
    Any other methods have you seen?

    Thank you.

  2. #2
    Join Date
    Feb 2013
    Posts
    602
    Hello,

    To make it simple :

    Your traffic is measured the whole month, in 5 minutes intervals.

    So lets say, in first 5 minutes you are at 110 Mbps, then in the next you are in 80, after 5 minutes, you are at 150 Mbps. (Doesn't matter if it is inbound or outbound traffic).

    In the end of the month, they will remove the 'top' 5% peaks, from all those measurements, and you'll be charged on the highest measurement left.

    There aren't other methods, the base of 95th percentile is exactly this, the only difference is when you choose to don't charge inbound or outbound traffic, but this depend on your traffic usage strategy.
    Mark
    IT / software / infrastructure / security projectist.
    Specializing in datacenter environments and threat mitigation.
    Your success, is our success !

  3. #3
    Join Date
    Dec 2009
    Posts
    88
    I know that not everyone uses the same sample rate. Would it make much of a difference between 1min/5min/10min?

  4. #4
    Join Date
    Feb 2013
    Posts
    602
    Quote Originally Posted by cyberiadmin View Post
    I know that not everyone uses the same sample rate. Would it make much of a difference between 1min/5min/10min?
    Doesn't matter the interval, as long you skip the highest 5% of measurements.
    Mark
    IT / software / infrastructure / security projectist.
    Specializing in datacenter environments and threat mitigation.
    Your success, is our success !

  5. #5
    Join Date
    Aug 2000
    Location
    Sheffield, South Yorks
    Posts
    3,480
    Quote Originally Posted by MAXZEUS View Post
    Doesn't matter the interval, as long you skip the highest 5% of measurements.
    Actually the interval does matter, traffic sampled at 1 minute intervals will be spikier and thus generally have a higher 95th percentile reading than traffic sampled at 5 minute intervals.

    Lets so for 30s I do 100Mbit/s then nothing for the rest of a 5 minute period - I do this for 50 minutes.

    - With 1 minute sampling the first sample will be 50Mbit/s the rest 0Mbit/s at the end of the period I have 10 readings of 50Mbit/s and 40 of 0Mbit/s (Assuming the 30s of traffic falls within one of the 1 minute intervals and not over two of them).

    - For 5 minute sampling each will be 10Mbit/s giving me 5 samples of 10Mbit/s.

    The 95th for 1 minute sampling will be 50Mbit/s, for 5 minute sampling the 95th will be 10Mbit/s. Extreme examples, but you can see the difference it makes.
    Karl Austin :: KDA Web Services Ltd.
    UK Business Hosting and Managed Servers - Hosting for Business Users :: 0800 5429 764
    Call us today and ask about our hosting solutions.

  6. #6
    Join Date
    Feb 2013
    Posts
    602
    Quote Originally Posted by KDAWebServices View Post
    Actually the interval does matter, traffic sampled at 1 minute intervals will be spikier and thus generally have a higher 95th percentile reading than traffic sampled at 5 minute intervals.

    Lets so for 30s I do 100Mbit/s then nothing for the rest of a 5 minute period - I do this for 50 minutes.

    - With 1 minute sampling the first sample will be 50Mbit/s the rest 0Mbit/s at the end of the period I have 10 readings of 50Mbit/s and 40 of 0Mbit/s (Assuming the 30s of traffic falls within one of the 1 minute intervals and not over two of them).

    - For 5 minute sampling each will be 10Mbit/s giving me 5 samples of 10Mbit/s.

    The 95th for 1 minute sampling will be 50Mbit/s, for 5 minute sampling the 95th will be 10Mbit/s. Extreme examples, but you can see the difference it makes.
    It will still be 95%, the only thing that will change is the 'precision' on the measurements, which can be better with lower intervals and less precise with higher intervals.

    However, the 95% is still there .
    Mark
    IT / software / infrastructure / security projectist.
    Specializing in datacenter environments and threat mitigation.
    Your success, is our success !

  7. #7
    Join Date
    Aug 2000
    Location
    Sheffield, South Yorks
    Posts
    3,480
    Quote Originally Posted by MAXZEUS View Post
    It will still be 95%, the only thing that will change is the 'precision' on the measurements, which can be better with lower intervals and less precise with higher intervals.

    However, the 95% is still there .
    Yes it's still a 95th percentile measurement. But I think given we're in the colocation forum, we're talking about bandwidth calculations then we can rightly assume when they say, "does it make a difference" they mean, "Does it make a difference to the number that comes out of the calculation". So to say that the interval doesn't make a difference when it makes a very fundamental one is not very helpful.
    Karl Austin :: KDA Web Services Ltd.
    UK Business Hosting and Managed Servers - Hosting for Business Users :: 0800 5429 764
    Call us today and ask about our hosting solutions.

  8. #8
    Join Date
    Feb 2013
    Posts
    602
    Quote Originally Posted by KDAWebServices View Post
    Yes it's still a 95th percentile measurement. But I think given we're in the colocation forum, we're talking about bandwidth calculations then we can rightly assume when they say, "does it make a difference" they mean, "Does it make a difference to the number that comes out of the calculation". So to say that the interval doesn't make a difference when it makes a very fundamental one is not very helpful.
    It depends on your point of view, for a 95th percentile measurement, the interval doesn't matter.

    If you ask about the precision, then it is another history.
    Mark
    IT / software / infrastructure / security projectist.
    Specializing in datacenter environments and threat mitigation.
    Your success, is our success !

  9. #9
    Join Date
    Oct 2005
    Location
    London, UK
    Posts
    138
    Quote Originally Posted by MAXZEUS View Post
    It depends on your point of view, for a 95th percentile measurement, the interval doesn't matter.

    If you ask about the precision, then it is another history.
    Interval matters a great deal. You will find, in general, that 1 minute samples result in a higher 95th percentile than 5 minute samples. Especially if your traffic is more "bursty"
    Goscomb Technologies Limited - www.goscomb.net / AS39326
    E: [email protected] P: +44 (0) 203 129 4400 F: +44 (0) 203 129 4410
    IP Transit :: Colocation :: Dedicated Servers :: Leased Lines :: DSL
    Registered in England and Wales No. 05672987 - VAT Registration No. 853 7954 80

  10. #10
    Join Date
    Feb 2013
    Posts
    602
    Quote Originally Posted by goscombtech View Post
    Interval matters a great deal. You will find, in general, that 1 minute samples result in a higher 95th percentile than 5 minute samples. Especially if your traffic is more "bursty"
    We are speaking about 'the method', not about the 'precision of the method'.

    Given the method, have a basic calculation where 5% of the highest measurements are taken off, and the highest left one is considered as total used bandwidth for the time allocation (generally a month).

    The sampling rate will not 'change the method'.

    It will change the precision of the method, 'the method remains the same'.

    So, I am not denying whether it will affect or not the precision of the measurement, instead I am just stating the method will not change due to that, so once again, there is no difference in 1 minute, 5 minutes or 10 minutes sampling.
    Mark
    IT / software / infrastructure / security projectist.
    Specializing in datacenter environments and threat mitigation.
    Your success, is our success !

  11. #11
    Join Date
    Apr 2009
    Posts
    1,143
    Give it up maxzeus. In Regards to the OP questions..KDA is spot on.

  12. #12
    Join Date
    Aug 2002
    Location
    Seattle
    Posts
    5,512
    Standard convention is 5 minutes. If anyone is using something different it needs to be clearly disclosed to the customer.

  13. #13
    Join Date
    Feb 2013
    Posts
    602
    Quote Originally Posted by IRCCo Jeff View Post
    Standard convention is 5 minutes. If anyone is using something different it needs to be clearly disclosed to the customer.
    This is a valid point for the 'method concept' .
    Mark
    IT / software / infrastructure / security projectist.
    Specializing in datacenter environments and threat mitigation.
    Your success, is our success !

  14. #14
    Join Date
    Aug 2000
    Location
    Sheffield, South Yorks
    Posts
    3,480
    The aim of "the method" is to come out at the end with a number with which to bill the customer with, therefore the sampling rate is integral to "the method".
    Karl Austin :: KDA Web Services Ltd.
    UK Business Hosting and Managed Servers - Hosting for Business Users :: 0800 5429 764
    Call us today and ask about our hosting solutions.

  15. #15
    Join Date
    Feb 2013
    Posts
    602
    Quote Originally Posted by KDAWebServices View Post
    The aim of "the method" is to come out at the end with a number with which to bill the customer with, therefore the sampling rate is integral to "the method".
    The method (the concept of the measurement) of 95th percentile, doesn't specify the sampling rate.

    So, while it is part or a parameter of the implementation, it is not integral to the method, but more likely how the implementator want it to be.

    So yes, given the wanted parameter for the measurement sampling you'll have a result or another, but again, it isn't attached as a 'rule' for the method itself. As Jeff said, there is a 'standard' for this approach, which commonly tend to be '5 minutes sampling'.
    Mark
    IT / software / infrastructure / security projectist.
    Specializing in datacenter environments and threat mitigation.
    Your success, is our success !

  16. #16
    Join Date
    Jun 2011
    Location
    Portsmouth, UK
    Posts
    244
    I agree 5 minute sampling is the normal, we do 1 and 5 minute (1 for sys admin and 5 for billing)

    Both systems give you a 36 hour window during which you're not billed.

    1 minute sampling will create higher peaks, however 1 minute sampling will also chop off higher peaks. 5 minute sampling wouldn't see those peaks at all.

    It'll depend on your traffic profile as to which works best for you, but 5 minutes is the accepted industry standard for billing.

    Assuming WHT let me post links, this is the same uplink

    http://imgur.com/NAjGF4W,YKkhWjW#0 1 minute samples
    http://imgur.com/NAjGF4W,YKkhWjW#1 5 minute samples

    In this case the 95th for both is ~20Mb but it's not unusual for the 1 minute 95th to be substantially different.
    ServerHouse, UK Datacentre with roof access, satcoms specialist

    http://www.serverhouse.co.uk

  17. #17
    Join Date
    Dec 2009
    Posts
    88

    95th percentile methods?

    Thanks for all the information.

Similar Threads

  1. 95th percentile
    By Spunkyasp in forum VPS Hosting
    Replies: 2
    Last Post: 06-14-2009, 04:34 PM
  2. 95th percentile
    By mchaley in forum Colocation and Data Centers
    Replies: 3
    Last Post: 12-04-2007, 03:59 PM
  3. 95th percentile
    By onestar in forum Dedicated Server
    Replies: 23
    Last Post: 12-31-2003, 11:39 PM
  4. 95th percentile?
    By rlynch in forum Running a Web Hosting Business
    Replies: 5
    Last Post: 10-24-2002, 07:01 PM
  5. 95th percentile ???
    By nik martin in forum Hosting Security and Technology
    Replies: 3
    Last Post: 04-03-2002, 12:19 PM

Posting Permissions

  • You may not post new threads
  • You may not post replies
  • You may not post attachments
  • You may not edit your posts
  •