dslreports logo
site
 
    All Forums Hot Topics Gallery
spc

spacer




how-to block ads


Search Topic:
uniqs
18
share rss forum feed


funchords
Hello
Premium,MVM
join:2001-03-11
Yarmouth Port, MA
kudos:6
reply to espaeth

Re: Well worth reading on why P2P causes problems

There is no difference between uploading or downloading in implementation...
Correct, but there is a rather physical difference. When you are uploading, you are sending fat packets back-to-back. When you are downloading, you are sending tiny packets with large gaps between. Given any random moment when congestion occurs, it's the uploader that is very likely to lose a packet and initiate congestion control responses. The downloader avoids losing a packet simply because his tiny "ACK" packets present a smaller target.

TCP's congestion avoidance algorithm tunes itself based on round trip time calculations and packet loss. Since every TCP connection follows the exact same rules on how to deal with congestion, each TCP session is essentially equal on the network.
Only if each route is the same and only if each end-system is as responsive as all of the others. IRL, this is never the case.

The bottom line is that in the event of congestion, every TCP session backs off in roughly the same proportional amount.
If and only if all connections drop a packet (or they all provide whatever else congestion cue might be available on that route).

The people with the most TCP sessions pumping their data are going to win for throughput.
Not withstanding my arguments above, they are only going to win during that brief interval of recovery. If the network is not congested, there is no problem that Bob Briscoe's proposal fixes. It does not avoid congestion. It does not reduce the amount of bandwidth used by P2P.

It's a bit like saying -- when IRQ 3 goes up, it takes the CPU 300% more time to service it does to service IRQ 4 or IRQ 7. Therefore, when the system starts to crawl, we're going to triple the CPU time given to IRQ 4 or IRQ 7 (or we're going to skip IRQ 3 66% of the time -- just to be fair). At that point, is the real problem fairness?

And to you I ask -- should the network be running at "congestion" often enough and for periods long enough for Bob Briscoe's suggestion to actually matter very much? Is a network that runs at "congestion" for long durations a healthy one?

And to everyone I ask -- when you read George Ou's article, is his intent to solve a problem? Or is George really trying to brand BitTorrent as some kind of an exploiter -- so that any thoughts about Network Neutrality would only apply fractionally toward BitTorrent based on how many open TCP connections it has (including the idle connections)?

The algorithm is remarkably fair for bandwidth per TCP session, where the whole thing gets hosed up is the number of TCP sessions per end user.
Which only comes into play when the network capacity is exceeded -- a condition which network providers are expected to avoid. We pay ISPs to meet user demand, not to shape it into their own vision of what user demand ought to be.
--
Robb Topolski -= funchords.com =- Hillsboro, Oregon
"We don't throttle any traffic," -Charlie Douglas, Comcast spokesman, on this report.


justbits
More fiber than ATT can handle
Premium
join:2003-01-08
Chicago, IL
Reviews:
·Comcast Business..
said by funchords:

And to everyone I ask -- when you read George Ou's article, is his intent to solve a problem? Or is George really trying to brand BitTorrent as some kind of an exploiter -- so that any thoughts about Network Neutrality would only apply fractionally toward BitTorrent based on how many open TCP connections it has (including the idle connections)?

The algorithm is remarkably fair for bandwidth per TCP session, where the whole thing gets hosed up is the number of TCP sessions per end user.
Which only comes into play when the network capacity is exceeded -- a condition which network providers are expected to avoid. We pay ISPs to meet user demand, not to shape it into their own vision of what user demand ought to be.
You've singled out BitTorrent, when it's not the only P2P protocol that's out there. The graphs, in particular, show several other P2P protocols. Modified BitTorrent clients and other P2P apps are designed to be the most greedy of all network users.

Yes, we ideally are paying ISPs to provide quality Internet service, but does your Terms Of Service Agreement say anything about quality of service? Most likely _no_. Yes, they should have been and should be upgrading their pipes continually. Yes, they likely have been shucking their responsibility on this in favor of providing returns to investors. So, if a change to everybody's TCP stack would make it less necessary for ISPs to spend money on network traffic control devices like Sandvine and instead spend money on upgrading the network, I'm all for that! The problem that Sandvine excessively solves ideally results in a better network experience for all users. So, modifying all TCP stacks to have better congestion control could result in less need for network traffic devices or policies that are as aggressive as Sandvine's.

So, if the network traffic becomes more self-policing/self-regulating by modifying underlying existing Internet protocols, then, maybe a paradigm shift will occur. Maybe then network neutrality will become a battle ground that's more directed at ISPs failing to provide bandwidth instead of it currently being an attack against ISPs for attempting to implement their own forms of congestion control.


funchords
Hello
Premium,MVM
join:2001-03-11
Yarmouth Port, MA
kudos:6
said by justbits:

You've singled out BitTorrent, when it's not the only P2P protocol that's out there. The graphs, in particular, show several other P2P protocols. Modified BitTorrent clients and other P2P apps are designed to be the most greedy of all network users.
I did this on purpose. Worldwide, BitTorrent is top dog by a long shot. Also, #2-#4 (DC++, ED2K, Gnutella/G2) behave more like FTP and HTTP -- and so they can't 'exploit' as George Ou describes. The 2006 chart is from Japan -- which not only has a very different in-country architecture than in North America, but also has a very different pattern of application adoption. North American ISPs don't want to show us their charts, even to researchers working under NDA.

said by justbits:

Yes, we ideally are paying ISPs to provide quality Internet service, but does your Terms Of Service Agreement say anything about quality of service? Most likely _no_.
Actually, now that you mention it, it does (-- in a way)! (And I had not thought of this before now, so if this idea is half-baked, it's because it's half-baked.)

The name of the Comcast tier that I am on is called 6Mbps. Six months ago, Verizon Wireless settled with the State of New York because they were advertising Unlimited service that was not, in fact, unlimited. Now a settlement sets no court precedent, and a New York case may only have consultative value outside of that state, but it does demonstrate that such a case would have merit.

That aside, your question has nothing to do with this topic. We're not talking about Quality of Service here (unless I missed your point)?

The problem that Sandvine excessively solves ideally results in a better network experience for all users.
All? Have you read my original report? Sandvine deteriorated my network experience (link in signature). That's how Comcast got caught! Comcast improved the experience for some at the expense of others -- even if those others were completely within the boundaries of the law and their subscription terms.

everybody's TCP stack would make it less necessary for ISPs to spend money on network traffic control devices like Sandvine and instead spend money on upgrading the network, I'm all for that!
Upgrading the network instead of wasting it on Sandvine is a simpler solution. The logic of your thinking reminds me of something from the "Bastard Operator from Hell." It goes something like this:

Help Desk: "You would like more space?"
Caller: "Yes, please!"
Help Desk (typing commands): RMDIR %USERPROFILE% /S /Q (command to delete user's files)
Caller: "AIEEEEEEEEEEEEEEEEEEEEEH!"
Help Desk: "You should have plenty of space now!"
Secondly -- and I said this above, too -- making the proposed changes does nothing to alleviate congestion. It only changes the behavior during the brief moments between congestion and recovery. Said differently, it doesn't prevent your computer from crashing, it just changes the order of the reboot process.
--
Robb Topolski -= funchords.com =- Hillsboro, Oregon
"We don't throttle any traffic," -Charlie Douglas, Comcast spokesman, on this report.


espaeth
Digital Plumber
Premium,MVM
join:2001-04-21
Minneapolis, MN
kudos:2

1 edit
said by funchords:

Upgrading the network instead of wasting it on Sandvine is a simpler solution.
You keep making this statement over and over again like it's an obvious binary decision. I've said this before, but that statement just keeps reading as "To solve your debt problem you should just acquire more money rather than wasting time cutting spending."

Given the state of where MSOs are at with frequency capacity and deployable DOCSIS technology (ie, still pre 3.0) the only way to "upgrade the network" is to do node splits or ungodly expensive node additions. (ie, poured concrete slab, utility power, cabinet, trench in fiber, split the coax plant, re-engineer the amp placement and gain structure, etc)

You represent the options as cost equivalent, but they're not even in the same ball park. For the cost of a single node addition Comcast can probably buy Sandvine appliances for an entire region. You're talking about a box with a couple power cords, a couple network connections, and some licensing overhead compared to right-of-way contracts, conduit & fiber, power utility installation, cabinet/equipment fees, and a significant number of engineering hours to implement.

TV services are where the cable companies make most of their money. If it were TV-related services driving the need for expansion it might have a better chance at gaining funding. Since HSI is the only product driving the need for expansion, managing the traffic is the cheaper/faster/better approach for right now. I don't have a problem with MSOs practicing this type of network management as long as they are up-front about it. The reality is that some level of filtering will always be present on the Internet, the same way that our "free" society still has laws. Management will always be required to curb "abuse", be it mitigating Denial of Service attacks, filtering SPAM, or restricting network traffic in accordance with "Fair Access Policies". The big issue is that the rules and methods of traffic management must be disclosed; wishing for a complete "hands-off" approach to network management is a pipe dream that will rapidly turn into a nightmare.


funchords
Hello
Premium,MVM
join:2001-03-11
Yarmouth Port, MA
kudos:6
said by espaeth:

You represent the options as cost equivalent, but they're not even in the same ball park. For the cost of a single node addition Comcast can probably buy Sandvine appliances for an entire region. You're talking about a box with a couple power cords, a couple network connections, and some licensing overhead compared to right-of-way contracts, conduit & fiber, power utility installation, cabinet/equipment fees, and a significant number of engineering hours to implement.
The Sandvine products containing the P2P policy enforcement has been the cash-cow for Sandvine. Comcast is believed to be more than 50% of its business, according to numerous analysts. Sandvine is predicting its annual income to be $80-85 Million. How much does it cost to split a node or add a node? Nobody is talking.

Don't assume that just because it's a box with two wires that it is cheap. There is a significant investment in R&D in these products, and due to patents, Sandvine can exclusively offer certain of their technology. All this means that the cost of raw materials has nothing to do with the purchase price of the device.

And I'm not sure it will cost them much at all -- take this quote from Tony Werner, CTO of Comcast:
quote:
The return bandwidth is not on the worry list right now, for a bunch of reasons. For one, were splitting a lot of nodes based on the success of voice, high-speed Internet, and VOD. In other words, all based on downstream requirements, not upstream.

On HSD (high-speed data), Im using two to three 3.2 MHz carriers (upstream). A lot more than that are sitting fallow in my CMTS cards. In most markets, I still have 12 MHz of bandwidth I can reclaim from circuit switched voice, once we migrate off of those platforms. So for now, the 5-42 MHz to me seems plenty adequate.

...

As we hit 70 percent utilization, we issue a work order to split the node. But it depends on utilization. Usually we set it to split to 250 homes. And for us, 65 percent of our node splits are really decoupling of nodes at the headend."

»www.cedmagazine.com/how-sexy-is-···nty.aspx
--
Robb Topolski -= funchords.com =- Hillsboro, Oregon
"We don't throttle any traffic," -Charlie Douglas, Comcast spokesman, on this report.


espaeth
Digital Plumber
Premium,MVM
join:2001-04-21
Minneapolis, MN
kudos:2
said by funchords:

Comcast is believed to be more than 50% of its business, according to numerous analysts. Sandvine is predicting its annual income to be $80-85 Million.
Annual income in which calendar year? They can only sell the hardware once, and recurring subscription/maintenance fees aren't going to drive the same level of return as the initial capital outlay.

said by funchords:

How much does it cost to split a node or add a node? Nobody is talking.
Comcast talks to investors all the time -- I get a nice prospectus from them every year that gives a high level overview of the company, it's operation, and it's financial performance. There are a few different methods of splitting a node. Most nodes start out combined at the head-end with multiple HFC nodes sharing a common head-end port (at least on the downstream). Those can be split by simply breaking the nodes into their own CMTS ports at the head-end. According to Comcast investor numbers this costs about $7500 per split ($2500/year on a 3 year depreciation cycle). The next split happens at the node itself where there is typically a north, south, east, and west string that branches out from the platform. Each segment can be broken out to have its own dedicated frequencies back to the CMTS with dedicated fiber. The investor numbers that were given for that are $18,000 per split (again $6k/3years). If one segment of a node becomes heavy (ie, north) you need to inject another distribution point in the coax plant to divide the infrastructure out yet again. Comcast places a buildout value of $60,000 on that split ($20k / 3years). Personally I think that value is low, but it is possible if neighborhood platform buildout is not factored into the cost. (ie, a new development goes in and the developer establishes a concrete pad with power hookups for telco / MSO / utility equipment) As of the investor documentation Comcast sent out in 2007 they had approximately 102,000 HFC nodes on their network.

said by funchords:

Don't assume that just because it's a box with two wires that it is cheap. There is a significant investment in R&D in these products, and due to patents, Sandvine can exclusively offer certain of their technology.
My assumption wasn't based strictly on hardware, but on pricing that I know personally from having cut purchase orders. If I use boxes from F5 Networks and 8E6 Technologies as a starting point, my guess would be that each P2P appliance probably costs about $30k.

said by funchords:

And I'm not sure it will cost them much at all -- take this quote from Tony Werner, CTO of Comcast
You have to keep in mind the audience of that article -- you don't tell investors that you're backed into a corner when it comes to your infrastructure. As far as expandability goes, he's right there are a lot of options for long term growth but all that development takes time. It also glosses over strategic value of certain upgrades. For example, if you've got open ports on your DOCSIS 1.1 CMTS line cards then doing node splits is no big deal. When you start talking about procuring hardware things become a lot more sticky -- this gear either has a 3 or 5 year depreciation cycle, so with the company's aggressive stance on DOCSIS 3.0 deployment the last thing they want to do is sink a bunch of capital into pre-3.0 hardware and be stuck with it past the end of the decade.


funchords
Hello
Premium,MVM
join:2001-03-11
Yarmouth Port, MA
kudos:6
said by espaeth:

said by funchords:

How much does it cost to split a node or add a node? Nobody is talking.
Comcast talks to investors all the time -- I get a nice prospectus from them every year that gives a high level overview of the company, it's operation, and it's financial performance. There are a few different methods of splitting a node. Most nodes start out combined at the head-end with multiple HFC nodes sharing a common head-end port (at least on the downstream). Those can be split by simply breaking the nodes into their own CMTS ports at the head-end. According to Comcast investor numbers this costs about $7500 per split ($2500/year on a 3 year depreciation cycle). The next split happens at the node itself where there is typically a north, south, east, and west string that branches out from the platform. Each segment can be broken out to have its own dedicated frequencies back to the CMTS with dedicated fiber. The investor numbers that were given for that are $18,000 per split (again $6k/3years). If one segment of a node becomes heavy (ie, north) you need to inject another distribution point in the coax plant to divide the infrastructure out yet again. Comcast places a buildout value of $60,000 on that split ($20k / 3years). Personally I think that value is low, but it is possible if neighborhood platform buildout is not factored into the cost. (ie, a new development goes in and the developer establishes a concrete pad with power hookups for telco / MSO / utility equipment) As of the investor documentation Comcast sent out in 2007 they had approximately 102,000 HFC nodes on their network.
Awesome! I spent about a half-hour looking for this information in Google -- and I probably can find that prospectus on the Investor Relations site. :::SHEEESH::: (BTW, there's some interesting "physical node split" alternatives being offered -- at least interesting by title only -- when searching Google).

$60K strikes me as low, too. Quite honestly, the numbers in my head had 6 digits and they didn't start with a 1.

The depreciation thing confuses me -- does this mean they can't charge the entire expense in the same year? Or does it mean they can charge the entire expense in the purchase year and then write-off the value of the capital (as a loss) over the next 3 years? (Other than taxes, does the depreciation amount have any value to this debate?)

the last thing they want to do is sink a bunch of capital into pre-3.0 hardware and be stuck with it past the end of the decade.
If Verizon doesn't get back on the stick, perhaps they can.
--
Robb Topolski -= funchords.com =- Hillsboro, Oregon
"We don't throttle any traffic," -Charlie Douglas, Comcast spokesman, on this report.


funchords
Hello
Premium,MVM
join:2001-03-11
Yarmouth Port, MA
kudos:6
reply to espaeth
said by espaeth:

My assumption wasn't based strictly on hardware, but on pricing that I know personally from having cut purchase orders. If I use boxes from F5 Networks and 8E6 Technologies as a starting point, my guess would be that each P2P appliance probably costs about $30k.
So, based on numbers like yours, and the fact that 65% of Comcast's node splits are of the "virtual" kind (either the $7.5K or $18K type), please tell me if you think is it safe-ish to say: "Comcast spends about as much on node splits as they do on Sandvine."
--
Robb Topolski -= funchords.com =- Hillsboro, Oregon
"We don't throttle any traffic," -Charlie Douglas, Comcast spokesman, on this report.


espaeth
Digital Plumber
Premium,MVM
join:2001-04-21
Minneapolis, MN
kudos:2
How many Sandvine appliances do you honestly think they have? My guess would be no more than a pair per market head-end.


funchords
Hello
Premium,MVM
join:2001-03-11
Yarmouth Port, MA
kudos:6
Wow. I was thinking more than that -- some number higher than 1 per 10 Gbps hanging off the router at the aggregation point. (That's of the PTS 14000 model). I have no idea regarding the 8210s which they probably have, too.