Join Date
Aug 2002
Location
uk
Posts
163

## 95th percentile

Can someone explain this method of measurement please:

Ok I understand that readings of throughput ? are taken every 5 mins eg reading1=10mb/s

The top 5% measurements are dropped.

The results are then correlated to total Gb throughput.

Does that sum it up ?

My question is why not actual throughput ?
400Gb month = 1Mbs (approx) continously on but if nothing is limiting throughput then the measurements during those readings could be 30Mbs (for example) lower during 4 mins, peaking during the 5th minute (for arguments sake)

Is that right ?

So how does this become an accurate method of measuring throughput ?

Thanks

2. Web Hosting Master
Join Date
Jun 2000
Location
Southern California
Posts
12,121

## Re: 95th percentile

Originally posted by onestar
...but if nothing is limiting throughput then the measurements during those readings could be 30Mbs (for example) lower during 4 mins, peaking during the 5th minute (for arguments sake)

So how does this become an accurate method of measuring throughput ?
You are correct. If your site is hitting 30Mbps, then this is your usage, this is what your site requires, and thus this is what would be factored in towards your bill with 95th percentile billing.

Your provider has to have the capacity to handle your peaks. Even if their total average is 1Mbps, or the total amount that is transferred equals 400Mbs, your provider must have the capacity to serve far beyond that. Beyond your peak levels even so that they don't top out at (using your example) 30Mbps and things become slow at that point.

3. Web Hosting Master
Join Date
Jul 2002
Posts
924
Some vacation chicken...

Join Date
Aug 2002
Location
uk
Posts
163
Ok understand the reply, and I think I know what I meant to ask.

Numbers chosen for example:

If for 2secs someone downloaded a file, the reading was taken at that point it could possibly be 30mbs for say a 5 megbyte file. eg 40mbits, 1 sec to download, answer 1.2 secs and reading taken at that point at 30mbs. Is this correct ?

Assuming that - then in the final tally this 30mbs reading is taken as the reading for the 5 min period (correct ?) eg correlates to 1 gig transferred.

Now I know we're talking stats, but is it possible that this happens 900 times in the month each time occuring on the reading point. The tally shows 900G transferred, when infact only 4.5 Gb was transferred.

This is my question about accuracy. I understand the bursting speed, but if the charge is per GB transferred could this not pose a problem ?

Thanks

5. Newbie
Join Date
May 2002
Location
Posts
11
Example:

12:05 PM - At this point 250MB has been transferred since the interface was brought up.

12:10 PM - At this point 350MB has been transferred since the interface was brought up.

When calculating the throughput for that 5 min period, you subtract the MB and times... so

Throughput = (350 - 250) / 5 mins = 100MB/300 secs = 0.33MB/sec = 2.67mbps.

It doesn't matter if all 100MB were transferred in the last 1 second or not...

Join Date
Aug 2002
Location
uk
Posts
163
Is thats what is hapenning ? That makes more sense,I did actually think they were taking snap readings of throughtput at that moment of time.

Since they are taking actual readings why convert in mbs then ?

Join Date
Aug 2002
Location
Beautiful suburb of Chicago
Posts
141
Someone charging by GB/mo should be counting and measuring packets, not bandwidth. Conversions should not be an accepted practice for billing. We have clients moving 140 GB/mo with a 95th measurement of 1.8 Mbps, and we have clients moving 180 GB/mo with a 95th measurement of 0.8 Mbps. Both are possible, but one should be billed for quite a bit more than the other.

If you feel that the above explanations aren't detailed enough or lack examples (I personally don't), let me know, and I'll write up an essay for you. =)

George

Join Date
Aug 2002
Location
uk
Posts
163
Well if my understanding is correct then I'm okay.

The 5 minute reading is the switch reporting back the average rate during the five minute period -- not taking point A and point B and dividing by 2.

Or it is actual readings of packets transferred during those 5 minutes.

Would that be correct?

But George you showed two examples of 1.8mbs and 0.8mbs how would that happen if the above is correct?

9. Web Hosting Master
Join Date
Jun 2002
Location
PA, USA
Posts
5,137
I once wrote the following, in an attempt to explain what 95th percentile is. From my own experience, measuring bandwith by 95th percentile (max of either incoming or outgoing) is roughly about 1.6 times that of measuring bandwith by average (sum of both incoming and outgoing). Of course, if you know how to maximize your incoming bandwith (minimizing the difference between incoming and outgoing) or if your bandwith patterns are close to being the same around the clock, then I will personally chose 95th percentile over average/actual.

95th percentile has always raised a question to most people. I will try to explain it the best of my knowledge:

To measure bandwith, one can use MRTG to measure the data transfer rate for an interval of every 5 minutes (it can be 30 minutes or 2 hours interval, but I just choose 5 minutes for this explanation). So for example, during the first 5 minutes, the server push 100MB of data, MRTG will record this as xxx kbps of bandwith transfer rate for that 5 minutes interval. Now there are 8640 samples of these 5 minutes intervals for a 30-day month. The way 95th percentile works is it will sort all of these 8640 samples, from smallest to the highest and then cut the top 5% of these 8640 samples (or the top 432 sampels). The next number (the 8207 th sample) will be your 95th percentile measurement of your bandwith usage.

The MRTG of course will measure not only incoming but ougoing transfer so there are actually 8640 samples of incoming transfer rate and another 8640 samples of outgoing transfer rate. By applying the above rule, you can get 95th percentile measurement of your incoming and outgoing bandwith. But then you will only be charged by the biggest of the two, either incoming or outgoing. This is the reason why if you can maximize your ration between incoming and ougoing bandwith, 95th percentile can be very cheap. Note with actual data transfer or average bandwith measurment, you will always be charged the sum of your incoming and bandwith transfer.

Join Date
Aug 2002
Location
Beautiful suburb of Chicago
Posts
141
The discrepency arises because of the measurement intervals and the burstiness/filesize in question.

A lot can happen in 5 minutes that doesn't get recorded by MRTG. Pushing 1.5 Mbps (DSL, for example) for 4 minutes, that's 360 Megabits, which is 45 Megabytes, or 0.045 GB, or 13.5 cents at \$3/GB. Granted that one single event such as this is just a dime and couple of pennies, but consider this: it can (and does) happen all the time. For this reason, many upstream carriers are now measuring in 10, 15, 30 second intervals. It means greater revenue for them because fewer packets squeeze through the smaller measurement windows.

George

11. Web Hosting Master
Join Date
Aug 2002
Location
Kelowna B.C.
Posts
1,687
This is the reason why if you can maximize your ration between incoming and ougoing bandwith, 95th percentile can be very cheap.
Yes true, but 95th percentile providers always count on you to use way more download than upload.

12. Web Hosting Master
Join Date
Jun 2002
Location
PA, USA
Posts
5,137
Originally posted by Hostkookster
Yes true, but 95th percentile providers always count on you to use way more download than upload.
Which providers will this be? My contract says "either incoming or outgoing".

13. Community Guide
Join Date
Jun 2000
Location
Washington, USA
Posts
5,991
It's usually calculated based on whatever direction used the most bandwidth. Because bandwidth is purchased in a symetric manner.

14. Disabled
Join Date
Feb 2003
Location
NY
Posts
586
servercentral gives free incoming b/w

15. Disabled
Join Date
Feb 2003
Location
NY
Posts
586
doesnt incoming b/w help with peering arrangements ?

16. Disabled
Join Date
Feb 2003
Location
NY
Posts
586
Do you guys know any hosts who bill on 95% of the SUM of both incoming and outgoing?

17. Web Hosting Master
Join Date
Nov 2002
Posts
2,780
Yes, I know it. It's lame to bill incoming + outgoing, but I seen it done.

18. Disabled
Join Date
Feb 2003
Location
NY
Posts
586
Yea I also saw someone doing it. Doesnt make any sense

19. Web Hosting Master
Join Date
Jun 2000
Location
Southern California
Posts
12,121
Note: thread posted 08-23-2002 in case anyone was going to reply to any of the discussion

20. Disabled
Join Date
Feb 2003
Location
NY
Posts
586
yea and ? I did a search on it and had a relevent question some ppl do search before posting

21. Web Hosting Master
Join Date
Jun 2000
Location
Southern California
Posts
12,121
Sorry, I meant besides your questions (I should have made that clear). And you searching before posting just makes you a freak of the board (I'm kidding you of course...)

RE: "Do you guys know any hosts who bill on 95% of the SUM of both incoming and outgoing?"

Someone posted about this on HHO. They weren't particularly pleased about it (can't blame 'em). I'm not sure if they named names or not (can't recall), and in the end they moved away from there as a result of the discussion.

22. Disabled
Join Date
Feb 2003
Location
NY
Posts
586
mfjp who were you talking about btw ?

23. Disabled
Join Date
Feb 2003
Location
NY
Posts
586
chicken I cant find that thread on hho can you link me ?

Thanx

24. Web Hosting Master
Join Date
Jun 2000
Location
Southern California
Posts
12,121

In this post he names 'em. he also mentions a Google link -but as far as I could tell, names were not named in that thread, not sure if it's the same company or someone else. You'd have to contact the thread starter there.

You can probably get away with doing it, but it's not the norm/typical and Brian's biggest beef was that, "I can't compare apples to apples with other providers." That's certainly understandable, even besides the higher amount charged.

#### Posting Permissions

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