Ubuntu Download Speeds Beat Windows XP's 515
narramissic writes "Doing a download speed test of his Time Warner cable connection, James Gaskin discovered something odd, something that he is quick to note isn't a rigorous benchmarked lab test. The discovery: His Ubuntu machine 'returned a rating from the Bandwidth.com test of 22-25mbps over several tests' while the same test done from a Windows XP PC returned a rating of 12-14mbps. The two computers used in the test are 'almost identical: both off-lease Compaq small form factor D515s, part of the very popular corporate desktop D500 family. Both have Pentium 4 processors running at 2GHz. The Ubuntu machine has 768MB of RAM, while the XP box has only 512MB of RAM. Both run Firefox 3 as their browser.' Gaskin's question: Can a little extra RAM make that much difference in Internet download speeds or does Ubuntu handles networking that much faster than Windows XP?"
TCP/IP Optimization (Score:5, Informative)
I'd guess it's some kind of TCP/IP optimization (the default size of packets, etc). It's set to one thing on Ubuntu, and another on Windows (probably for some historical reason or due to some old buggy driver).
If that's not it, I'd bet pretty high it's a bad driver in Windows.
It's quite likely that either Windows or Ubuntu is intrinsically faster for some reason, but I doubt the difference based on the way the networking stack is designed is anywhere near 10%, let alone 50% for a link this fast. On 10 gigE maybe, but not on a simple cable modem.
this made it on slashdot? (Score:4, Informative)
So how did this make it to slashdot. Its not like anyone but the poster has the identical hardware to run the tests properly.
@poster: If the machines are so 'identical' then swap the memory and run the tests again.
Scientifically Bollocks (Score:5, Informative)
You can't test two different machines with different cases and compare the results, that's not how the scientific process works. Both machines need to be tested against the same cases - then and only then will you be able to appropriately tell if the software made a difference.
Anyhow, back on the subject, some of WinXP's default networking parameters are a bit conservative when it comes to high-bandwidth links that don't have LAN-like latency (particularly the TCP Receive Window/RWIN); a good but short description of this can be found at DSL Reports [dslreports.com]. So I wouldn't be absolutely shocked if once he corrects his methodology, he still gets similar results, although in general I find RWIN tweaking to be bollocks compared to the few people that swear it works. Vista and later OSs include self-adjusting network stacks that compensate for this and then some (Microsoft is rather proud of their sustained bandwidth over very high latency links), so I wouldn't lose any sleep over it.
Re:Even if the answer is no... (Score:5, Informative)
True, but considering both computers should easily be able to saturate a 100baseT connection, shouldn't both configurations be able to saturate a 22Mbps link?
This is different than the guy complaining that the computers can't fill a gigabit ethernet connection with a scp transfer while music is playing.
The http that the speed test should be using doesn't have any encryption, shouldn't be using gzip, and it shouldn't be saved to hard drive.
TCP Window Size is the likely culpret. (Score:5, Informative)
It's the TCP implementations, and probably the TCP window size limits. Windows could turn in the same numbers if properly tuned.
You want to read this article for all the in-depth details: http://www.psc.edu/networking/projects/tcptune/ [psc.edu]
Windows has a default set many years ago, and never updated. Most of the Free Unix variants update every release, and some new variants even have fancy auto-scaling code. Any time you want to get over 10Mbps/second across any real latency with a SINGLE TCP stream you probably need to do some tuning, for some OS's the limit is much lower.
ISP's run into this all the time. An uninformed admin buys a GigE in LA and NY, pops up an FTP server and wonders why he can only get a few megabits a second across the "crappy network". A few settings later and behold, the same hardware can saturate a full gigabit.
Note, don't just go set your values really high, there are performance (memory used) tradeoffs....
Re:No, haven't RTFA, thank you very much (Score:5, Informative)
Actually I can see it all the time. My Ubuntu laptop's (IBM T42) WiFi is about 50% faster that the same configured Windows machine of my wife. We're talking about SAME hardware. I don't really know if it's drivers, or something else.
Performance on LAN is more similar, difference is about 10-20% max, but with this kind of hardware it heavily depends on HDD to write data and Windows is crap at this - it's swapping - god knows why!!!
Re:amazing (Score:4, Informative)
BS. XP runs fine on 0.5GB ram. Hell, when it came out, what was the norm for a new machine? 128MB? 256?
You're thinking of Vista.
Driver, OS, Hardware - In That Order (Score:1, Informative)
Having done research about this for years in financial services labs, this isn't "odd" to me, nor has it been "odd" to me for about a decade. A lot of good hardware is hampered by rather lousy hardware driving writing and integration with the OS. Further, the Windows TCP/IP stack hasn't been known to be a hugely big performer on the desktop side.
A TCP/IP stack at its core is really about the queuing mathematics in the code itself. When you get a good match of that math between two systems, things go smoothly. When you don't, performance suffers.
Microsoft would do well to see the strides that Linux, but more so the FreeBSD folks have done in making TCP/IP stacks more efficient - hats off to the folks that have done TCP/IP stack research for better efficiency in the data center.
I'd also proffer the opinion that the threading / process code base in most UNIX OS's is vastly superior to the Windows model for efficiency - this isn't so much a comment about Microsoft as it is trade-offs made over time / different ideas.
To be fair, I've had good success in tweaking Windows TCP/IP stack parameters, but this is frowned upon in large enterprises because no one likes rolling out registry changes. Even with the tweaks, the raw network performance has always fallen short of a sample of Microsoft OS offerings. Microsoft might want to loosen things up a bit.
Re:Is this.... (Score:5, Informative)
That's a myth.
Clarification about the use of QoS in end computers that are running Windows XP
As in Windows 2000, programs can take advantage of QoS through the QoS APIs in Windows XP. One hundred percent of the network bandwidth is available to be shared by all programs unless a program specifically requests priority bandwidth. This "reserved" bandwidth is still available to other programs unless the requesting program is sending data. By default, programs can reserve up to an aggregate bandwidth of 20 percent of the underlying link speed on each interface on an end computer. If the program that reserved the bandwidth is not sending sufficient data to use it, the unused part of the reserved bandwidth is available for other data flows on the same host.
http://support.microsoft.com/kb/316666 [microsoft.com]
Re:amazing (Score:3, Informative)
The minimum spec for the original XP and for SP3 both is a 233Mhz processor and 64MB of RAM. A 300Mhz processor and 128MB of RAM was recommended. These were extremely low-balled numbers, but a system configured such would boot and run.
Many of the applications require much more than that, though. IE7 requires 64MB minimum for just itself. Here's that requirements page [microsoft.com].
If you take 64 MB for the OS and 64 MB for the browser, a 128MB system will probably swap from a single browser window loading a complex page, let alone doing a large download.
Now, add in Windows Firewall, some anti-virus software, and a couple of other resident programs. For testing, most of this should be turned off. The Windows Firewall I'd leave on because my Linux box would have iptables and possibly Shorewall or some other management wrapper around iptables running.
Firefox 3 isn't exactly stingy on memory use, if that's what he's using on both platforms. Neither is Flash, as it seems most download speed test web sites use.
So, yeah, he might be swapping pretty heavily at 512 MB although you're right that the base system would run okay with even less than that.
That's not the only explanation for such a difference, though. He might be running on-demand virus scanning against the download. He might not be telling us that he's saving the download to disk and one has a faster, after-market hard drive in it. An uncontrolled test is only of anecdotal value.
Re:TCP/IP Optimization (Score:3, Informative)
Its common knowledge that Windows has an inefficient TCP stack as far as higher speed broadband connections go.
Unblocka [unblocka.com] and TCP Optimizer [speedguide.net] are two apps commonly mentioned on the Australian Whirlpool [whirlpool.net.au] forums.
Re:Even if the answer is no... (Score:3, Informative)
Re:Even if the answer is no... (Score:1, Informative)
Who gets 22mbps from a cable modem?
Here in Helsinki, Finland people can easily get 50 mbps (advertised speed is 100 mbps) using the de-facto cable monopoly, Welho, and their (quite recently introduced) wideband cable modem service.
Moderators... (Score:3, Informative)
Same with databases (Score:3, Informative)
We used to regularly benchmark Oracle on the same hardware running Linux and then Windows Server. Linux always won. Not by a huge margin, more like 15%, but saving money and getting better performance is win-win!
Re:Even if the answer is no... (Score:5, Informative)
A new low in slashdot history. Can't remember a worse headline and article in recent times, can anyone else?
This article is just one big WTF. Is slashdot that desperate for traffic?
By these standards your nut article will indeed make a headline soon.
And why did my submission not get posted, yet:
My windows PC with a 27 inch screen runs at 1600x1200 resolution, my ubuntu on a 15 incher only 1024x768. Are windows graphics drivers better than the linux kernel?"
Re:TCP packet size. tcp window scaling. (Score:3, Informative)
The problem is usually that they decided to be extra careful about security and bought a PIX.
You'd think Cisco knew something about networking, but that knowledge certainly hasn't made it to the PIX/ASA department.
TCP Window Scaling (Score:3, Informative)
I believe XP has tcp window scaling turned off by default, whereas modern Linux kernels and Vista have it turned on.
See http://en.wikipedia.org/wiki/TCP_window_scale_option [wikipedia.org]
This can make a massive difference if there is more than a tiny amount of latency on the line...
Re:TCP packet size. tcp window scaling. (Score:5, Informative)
This is due to TCP Window Scaling (Score:3, Informative)
Window scaling is disabled by default on windows, which limits TCP sessions to 64 kB, hence the per-session bandwidth on high-latency links such as DSL.
10-12 Mbps is typical of a DSL link with a 50 ms RTT (=ping time). 64 kB is 512 kbit. 512 kbit / 0.050 s = 10240 kbps = 10 Mbps.
I've already seen tuning guides on the net explaining how to enable window scaling on windows, though I'm not that much interested ;-)
Willy
Re:Temporary boost in download speed? (Score:4, Informative)
I'd mod you up, but I feel compelled to reply ... since I'm amazed nobody has mentioned this.
I just signed up for Time Warner 'net myself, and when the dude was checking the signal he mentioned something about how there's a 25Mbit "boost" that people get at random. I didn't get a chance to ask many questions about it, but he said that it wasn't just an ISP-level cache ... you're actually given 25Mbit of bandwidth for a breif amount of time. That could very well be what we're seeing here, as the numbers seem to align.
Re:Linux on the desktop (Score:5, Informative)
Here's my totally unscientific test from my home office. Vista 32bit vs. Ubuntu 32bit. Done on a dual-boot Thinkpad T61 laptop, 2.0ghz 2gb RAM. Nowhere close to running out of memory, only app running was the browser (plus the normal tray stuff).
I did five tests with each OS/browser and averaged them out, doing the bandwidth.com test.
Figures in kbps. ISP is Comcast cable.
Windows Vista Chrome 1.0.154.43
Down: 18276.6 (avg) / 21522 (max)
Up: 1866.8 (avg) / 1898 (max)
Windows Vista Firefox 3.05
Down: 17357 (avg) / 23820 (max)
Up: 1044.6 (avg) / 1067 (max)
Ubuntu 8.10 (Intrepid) Firefox 3.05
Down: 15451.6 (avg) / 21742 (max)
Up: 2035.6 (avg) / 2151 (max)
The averages differed wildly but I think network traffic can easily account for this. Since the maximums were all nearly the same I think they're all about the same.
What it doesn't account for is the upload speed, which were very consistent throughout this silly test.
Vista firefox = dismal
Vista chrome = much faster
Ubuntu firefox = even faster
Re:-1, Hoary old joke (Score:3, Informative)
>I wish apple would sell a powerbook with a real right-click.
And I wish other laptops had the "two finger" right click and the two-finger scroll.
A real test (Score:1, Informative)
did he turn QoS off in WinXP? (Score:2, Informative)
This shouldn't be news!
Linux kernel has TCP auto tuning (Score:4, Informative)
Re:Linux on the desktop (Score:3, Informative)
Hey, where's the button? The amazing new trackpad doubles as a button â" just press down anywhere and consider it clicked. No separate button means there's 39 percent more room for your fingers to move on the silky glass surface. Now that Multi-Touch gestures have come to MacBook, all the function is in your fingers. Use two fingers to scroll up and down a page. Pinch to zoom in and out. Swipe with three fingers to flip through your photo libraries. Rotate to adjust an image with your fingertips. Using the new four-finger swipe gesture, swipe up or down to access Expose modes and left or right to switch between open applications. If you're coming from a right-click world, you can right-click with two fingers or configure a right-click area on the trackpad. The more you use the Multi-Touch trackpad, the more you'll wonder what you ever did without it.
Re:Linux on the desktop (Score:3, Informative)
whoosh.
I think the GP was talking about reaching orgasm.
Re:Linux on the desktop (Score:4, Informative)
Re:Linux on the desktop (Score:2, Informative)
Yes, for being too lazy to use LFS!
That's my experience, anyway. I was too lazy to fuck around with jhalfs and said fuckit after compiling the toolchain.
Gentoo, on the other hand...
Stage3 + emerge -e world + emerge -e world yields similar results to Stage1 + USE="" Stage2 + USE="" Stage3 + emerge -e world
With a prebuilt Stage3, I only have to do USE="" Stage3 + emerge -e world. Faster :D
YMMV.
The Tweak for vista... (Score:3, Informative)
ymmv this might mess up some sound during heavy downloads.. so far I havent had a problem.
Storm
Re:Linux on the desktop (Score:4, Informative)
They are talking about the Ancilliary Function Driver that is used for Windows sockets. By setting HKLM\CurrentControlSet\Services\AFD\DefaultSendWindow to 0x7800 allows them to set this to 480KB of data before flow control kicks in. Great if your hardware can handle it, not so great if you have crappy hardware. I believe that if you have > 32MB of RAM that the default is 8KB of data gets received before flow control throttles the connection.
Probably will work OK if your NIC can handle it and you have enough memory. And of course if you have high enough bandwidth :-)
Re:Ubuntu a zealous web hog? (Score:3, Informative)
The usual cause of this is programs which upload as fast as they can, saturating the upsteam link. Some programs are smart enough to guess how much upstream bandwidth you have and confine themselves to fit inside it without causing issues.
It makes normal usage slow because your computer sends acknowledgements as it receives data. If the host sending you that data doesn't receive the acknowledgements, it assumes there's congestion and reduces the rate at which it sends to you. If you're uploading at full steam, a large queue of packets will build up waiting to be sent out. The acknowledgement packets, while very small, also wait in this queue; and unless you're doing some kind of traffic shaping, it's just a simple first-in, first-out queue. The acks are therefore delayed for up to a few seconds before they get out.
The result is really slow browsing, as all of the latency sensitive stuff like DNS lookups and connection handshakes can take a very long time to complete. It can also slow down your downloads, since the rate at which the host you're downloading from will send you data is partially determined by how quickly (and consistently) you send acknowledgements.
The solution is either to use a client which magically determines your available bandwidth and reduces it upload rate so as not to allow queues to develop, or to manually tell your client to only upload at a certain rate. A more robust approach is to use traffic shaping techniques, which basically means putting a smart queue on the router so acknowledgements, DNS lookups and other small packets can skip ahead of the bulky upload packets in the transmit queue.
Re:Even if the answer is no... (Score:3, Informative)
Well, you may not have ATM but since ADSL's are mostly managed by large telco's, the chances are pretty big. My ADSL (and all other ADSL in the Netherlands as far as I know) uses PPPoA, which uses ATM, which uses AAL5, which uses cell switching using cells with 48 byte payload and 5 byte overhead (there was an argument between a group that wanted 32 bytes and one that wanted 64 bytes so they compromised).
More info here:
http://pflog.net/dsl_overhead/ [pflog.net]