Back in September I noticed that the speed when browsing the web in the Paris Metro with OperaMini has significantly improved using France Telecom's / Orange network. At the time I was wondering if this improvement over a previous massive slowdown was only temporary due to the vacation period or if something has really changed. This Saturday I took a hopelessly overcrowded Metro 4 through the city center and it's good to see that there was enough network capacity for my Opera Mini browser to fly from one page to the next. Thanks Orange, it's usable again now!
Almost exactly 2 years ago, I've first seen a demonstration of a cat Cat 7/8 HSDPA device and went away very impressed after seeing downlink speeds of 4.2 MBit/s. Now I've finally gotten my hands on one myself and tested it in Vodafone's network in Germany. On the 6th floor of a building, the base station pretty close by (I assume) and my download speeds easily exceeded 5 MBit/s (= 700 kbyte/s) consistently over several days and at different times of the day. Wow, that's twice as fast as my ADSL line. But that's not the end of the line, a few operators have already rolled out HSPA+ with 21 MBit/s. Examples are Mobilkom Austria in Vienna and Telstra in Australia.
A couple of weeks ago I posted my thoughts around doing throttling of heavy users directly at the base station so the current load can be taken into account as well. No need to throttle heavy users if nobody else is using the cell. While it was just a "Gedankenexperiment" a reader pointed my to this paper in which Ericsson actually describes just such an approach and says it has already implemented it in their RAN. Very interesting! The paper unfortunately doesn't contain a lot of information of how this is done from a technical point of view, i.e. how is the base station informed which users to prefer over others. If you know more, if this is not a state secret and some information about the feature is publicly available, please let me know!
Web surfing over 3G must be slower than over DSL, right? To see if this statement is correct, I used two computers side by side, one connected to to Internet over a 3 Mbit/s DSL line and the other one over a 3G dongle. The 3G network of choice was that of Vodafone Germany in Cologne with the base station at least supporting HSDPA category 8 (7.2 Mbit/s) devices.
For the first web browsing test I used a Huawei E220 3G dongle with a somewhat older but very reliable HSDPA category 6 software load. Computer screens side side by side I simultaneously clicked on links to load web pages, both visited and never visited before, to see on which computer the pages displayed first. The result: In this test, the web pages took around one second longer to be first displayed on the 3G connected computer but the difference was quite minor. Definitely less than what I expected.
In a second test, I used an HSDPA category 8 + HSUPA capable E176 which has been available on the market already for a little while. Definitely not the latest and greatest anymore but still good hardware. With this setup the side by side comparison showed no difference anymore, pages on both computers showed up almost instantly. Sometimes, the page would show on the DSL connected computer a fraction of a second earlier, sometimes it would be loaded a bit quicker over the 3G connection. Fantastic!
A little caveat: At the time of the test the network was only slightly loaded so one of these days I have to find a time at which it is a bit busier and repeat the test to see if that makes a big difference.
Back in 2006 I noticed that my Windows XP machine could not fully use the bandwidth of my ADSL line and also throughput over HSDPA to some servers was less than I expected. As I found out at the time, the fixed and small TCP window size was responsible for the behavior. In Windows XP, things could be tweaked by changing the window size in the registry as I described here. When running some throughput tests with Ubuntu Linux this week with an HSPA 7.2 MBit/s 3G stick, I noticed that no tweaking was necessary to get the full speed.
A quick look with Wireshark revealed why: Unlike Windows XP that has a static window size that is set somewhere around 17 kbytes, Ubuntu sets the TCP window size dynamically. It starts with a modest 5k window and steadily increases it during the file download to over 1 megabyte. Looks like Vista has a similar algorithm as well. Very nice, no more worries about throughput limitations in the future!
A little anecdote today: In the "old" days I had a 14.4 kbit/s fixed line modem for quite a number of years. Even though new and shiny 28.8 kbit/s modems came on the market, I was stuck with my '14.4' because the new modems were expensive and as a student my monetary resources forbade an upgrade. So for me, the number '14.4' has a bit of a negative touch attached to it ever since.
Fast forward to today to the "megabit" era. In wireless, HSPA 7.2 Mbit/s downlink is currently pretty much state of the art. Some network operators have announced further upgrades and in due time, top speeds of 21 MBit/s and beyond will be reached. On the way to double digit speeds, there's also a 14.4 step. No, not kbit/s, but Mbit/s. Still it kind of reminds me of my 14.4 kbit/s days and has a negative "psychological" touch to it to me.
Strange strange, because I'd really like to have this 14.4 this time around 🙂 Any numbers in telecoms that have a psychological edge for you?
When discussing High Speed Pack Uplink Access (HSUPA), or E-DCH, to use the correct term, the major focus usually lies on the improved uplink speeds. Seldom is it mentioned, however, that E-DCH also improves the latency, i.e. the time it takes for an IP packet to be sent to a server and a response packet to arrive back to the source. But is this relevant in practice?
So far, I used an HSDPA 3.6 non E-DCH capable 3G stick and my round trip delay times (RTD) to a number of web sites I visit most were around 110 milliseconds. Over a DSL link, the same sites can be reached with an RTD of around 45 ms. In other words, a difference of 55 ms. In practice this can be felt especially during web browsing, as web sites take a bit longer to load over 3G compared to a DSL link with a similar bandwidth. Not that this is a showstopper but it can definitely be felt.
A few days ago, I ran some tests with a Cat 8 HSDPA + HSUPA 3G stick and was quite surprised that the RTD times to those web sites were just around 65 ms. In other words, that's only 20 ms more compared to DSL. The difference to the HSDPA only 3G stick are quite remarkable. I compared backwards and forwards with my DSL line but I couldn't "feel" the difference anymore. Stunning!
The one thing E-DCH does not do away with, however, are the delays incurred when radio states are switched. The 300 ms or so delay when switching between a full DCH and the less power and resource intensive FACH are still there. In practice, however, background traffic from applications such as my Instant Messengers usually keep the link in DCH state so I rarely come into contact with it anyway.
Here's an interesting article from Ericsson on the business case of mobile broadband. Taking CAPEX, OPEX for both access and core network into account, the article comes to the conclusion that once an economy of scale is reached in terms of the number of broadband subscribers, the network can deliver 1GB of data for one Euro.
While this is the main outcome of the paper, there are a number of other pieces of information in there on which the calculation is based which are quite interesting. Here are some which I noted:
- 20% of the cells carry 50% of the traffic. I've heard of similar numbers before and I think it's a good thing because the network operator can focus on upgrading a subset of all cells rather than having to work on the whole network simultaneously.
- 3-5% of the cells carry very heavy load. The article doesn't say where such cells are usually located. It would be interesting if this load is mostly generated in-house, for example in shopping centers, train stations, airports, etc. and if femtos would provide a cheap future capacity extension for those places.
- The technical evolution of 3G networks is all about keeping pace with higher user demand for capacity. Fully agree to that.
- Going from 7.2 MBit/s to 21 MBit/s adds a cost of around 10-15% but increases capacity around 70%. An interesting statement because 7.2 -> 21 MBit/s is about a 3x theoretical speedup while from a practical point of view it is much less. The article says its 70% or 0.7x.
- 70% of overall CAPEX is spent on base stations.
- 50% of overall OPEX is spent on base stations.
- The €1 per GB seems to be a number over a 5 year period. At the end of the article it is stated that the networks that Ericsson looked at for the study are not quite there yet. However, the first network, after 2 years of operation, reported that they have reached €2 per GB.
- For the study, a base station price of €40k was assumed. Looks like they have gotten quite a bit cheaper than what was calculated with just a couple of years ago.
And just to get a bit more aggressive, this Ericsson presentation states that mobile broadband is even cheaper than DSL (cp slide 13 ff.). I assume that leaves IPTV delivery out of the equation, but still it's an interesting way of looking at things.
Almost exactly two years ago, I wrote a post in which I reported my first sighting of new HSPA+ device categories. The top at the time were category 14 and 16 with 64-QAM modulation and MIMO respectively and speeds up to 28 Mbit/s. This was Release 8 of the 3GPP standards. Now in Release 9, 28 device categories are listed in 3GPP TS 25.306 (see table 5.1a) with top speeds under ideal radio coverage of up to 84 Mbit/s if everything is combined, i.e. 64-QAM, MIMO and Dual Carrier. For almost every combination of options there's a category now. Breathtaking…
As I found no good overview of which device category goes up to which speed, I took the liberty of updating the HSDPA article on Wikipedia and add device classes 16-28 in the table. A screenshot of it can be seen on the left.
Now where can I get a Category 28 device and a suitable network please? 🙂
P.S. 1 – Important: Note that all indicated speeds are top speeds under ideal signal conditions. See here for further details and a reality check!
P.S. 2 – I left out cat 17 and 18 as they are a bit special and I am not sure that they will be relevant. If you have an opinion on this one, please let me know. Also, feel free to add them to the table on the Wiki yourself and while you are at it, have a go at the coding rates for the higher categories as well.
Most "all you can eat" wireless Internet access offer these days do come with a traffic limit per month after which the connection is either cut, the speed is decreased or further charges apply. The aim of these measures is to ensure that a few users do not disproportionally use the network. Especially in wireless networks, where air interface capacity is the limiting factor, network operators try to bring some fairness into the game. However, the throttling is not happening at the air interface but in the core network. For the sake of fairness, however, I wonder if that is the right place to do the throttling!?
Wouldn't it be better to to have a sort of a soft limit and control it via the base station traffic scheduler? Here, the scheduler could take into account how much data each user has already transmitted in the past hour or day and thus give the packets to or from this user a higher or lower priority. Doing this at the base station would have the advantage that while the cell is not loaded, even heavy users get the full bandwidth while under heavier load, users that only browse the web get a higher priority and are thus not significantly slowed down by streaming or downloading activities of other users.
The scheme doesn't work for moving users but I assume that most power users with a notebook use bandwidth hungry applications in a stationary mode. I am also aware that taking the user identity and past use into account at the base station scheduler is not standardized in 3GPP and I am not sure if the base station scheduler can keep track of a users identity over state changes (i.e. from Cell-DCH to Cell-FACH to Idle and back). Nevertheless, an interesting "Gedankenexperiment".
Oh yes, and by the way, this kind of soft-priorization is not new, it's done for satellite Internet connections already.