Monday 10 November 2008

Stimulacija za korisnike željezničkog prijevoza

Već dulje vrijeme razmišljam o tome kako usluge javnog prijevoza mogu postati popularnije među mlađom, informatički osvještenijom ekipom. Podjelit ću sa vama danas jednu od ideja.


WLAN internet na željezničkim kolodvorima, a možda jednom i u vlakovima u pokretu.

Evo opisa mogućeg rješenja: Svaki kupac željezničke karte ima pristup internetu u trajanju od 1-2 sata za vrijeme čekanja polaska vlaka na kolodvoru+ za vrijeme vožnje (dok vlak stoji na usputnim stanicama). U ovom slučaju svaki veći kolodvor ili stajalište vlaka ima nekoliko WLAN AP-ova, a softver u pozadini brine se da putnik može upisati broj sa karte kojim će se autorizirati u sustav. Korisnici pokaza i mjesečnih karata imali bi uvijek pravo na pristup, s time da bi se isto tako jednom autorizirali brojem na pokazu/mjesečnoj karti. Moguće prevare sa pokazima rješio bi sustav koji ne dozvoljava istovremeno korištenje na različitim kolodvorima, a uzima u obzir i MAC adrese i ukupno vrijeme korištenja pristupa internetu.

Ovime bi se postiglo više različitih stvari:

- putnici bi imali motiv doći ranije na vlak, te ranije kupiti kartu, i tako smanjiti gužve u prodaji karata, na ulasku u vlak i slično
- privukli bi se mladi zaposleni ljudi kojima se vrijeme stajanja na kolodvoru može pretvoriti u produktivni rad
- općenito bi ovakva akcija bila izvrstan marketinški potez za HŽ

Mogući partneri HŽ-u u ovakvom projektu bili bi: Optima Telekom, Metronet, Iskon, CARNet, a možda čak i T-Com.

Internet pristup za vrijeme vožnje bi imao smisla prvenstveno za vlakove koji voze na duge relacije, a mogao bi se izvesti HDSPA + WLAN kombinacijom, ili možda čak usmjerenim WLAN antenama na željezničkim postajama i prikladnim antenama na početku i kraju vlaka. Ili modulacijom frekvencije/napona elektrificiranih pruga kao što se radi u kućnim mrežama baziranim na strujnim vodovima. No nekako mi se ta zadnja opcija čini najskupljom :)

Saturday 5 April 2008

Radeon HD 3850 Power consumption/Linux

I had the opportunity to test a Radeon HD 3850 card recently and I've tested my two favorite hardware properties: Performance and power consumption. I won't go into too many details about the performance, it's sufficient to say that Crysis is playable with max details under 1024x768 with a Core 2 E8200 CPU. Of course Crysis won't work under Linux, at least not yet.

About the power consumption though:

I've recently bought a Kill-a-watt equivalent for the European market, an Art. Nr. 002580 from REV (description available at the German language page).



So I've measured the consumption of a complete system. 460W power supply from Coolermaster, P35-DS3P rev 2.0 motherboard from Gigabyte, 2 GB of Corsair CL5 RAM, and our subject the Sapphire Radeon HD 3850 Ultimate.

When the graphics driver was still not loaded (in the BIOS, during post, in the boot loader, and during the OS load) the machine was using 119 W. When Windows loaded power consumption under idle went down to 81 W. Under Linux idle consumption with the latest fglrx driver was 87 W, so somewhat higher. When I switched the VT to text mode, it went up to arround 119 W again. When the graphics driver under Linux does a LeaveVT to switch to text mode, it usually restores the state of the registers to the one found before the driver is loaded. That leads me to a conclusion that the video BIOS of the card doesn't do anything to save power, and we must wait for the driver to load for the PowerSaving features to be enabled. Regarding the open source "radeon" driver for Xorg, Alex Deucher just recently added DynamicClocks support, so that should help. I'll test it later and make an update to this post to let you know how it works out.

Here are some more values for the system if you're curious.

Crysis 1024x768 all details high: 153 W
3DMark 2001SE under Linux/Wine, Fill rate test, single texturing: 165-172 W (this is the champion)
3DMark 2001SE under Linux/Wine, High poly count: 123-128 W
glxgears Linux: 140 W

2x burnP6 (this test just exercises the CPU): 116 W

As a contrast the Thinkpad X60s that I'm writing this post on uses 38 W under maximum load, 17-19 W when idle, and around 22 W under typical use. Internal LCD panel consumption included in the figure of course. When in low power mode (typically on battery) it uses about 10-11 W. (Panel brightness lowered, wireless turned off).

Update on tha XAA/EXA saga

If you're using newer Xorg/mesa/intel driver, you will find that XAA doesn't work anymore. My machine tends to have a hard lockup. XAA probably broke because of the changes in the DRI/DRM 3D part of the driver.

Switching back to EXA makes it work, but with the same old performance problems. Still it's possible to make it go faster with a
Option "MigrationHeuristic" "greedy"
in the "Device" section of your xorg.conf


The mandatory sysprof of Firefox 3 doing business.hr test:



The exa manual page says this about the option:
Option "MigrationHeuristic" "anystr"
Chooses an alternate pixmap migration heuristic, for debugging purposes. The default is intended to be the best performing one for general use, though others may help with specific use cases. Available options include "always", "greedy", and "smart". Default: always.

Thursday 17 January 2008

UPDATE: Firefox/Mozilla performance under linux/Xorg

The addition of the Option "XAANoOffscreenPixmaps" "true" also helps with the Intel hardware I have. Be sure not to use EXA on the Intel because it also makes things really slow for now. Maybe EXA will be fixed some day on the "intel" driver, but currently it's better not to use it.

Here is the profile from the EXA intel current git (Firefox 2), it's slow:



Additional links:

business-exa-intel.sysprof.bz2

Firefox/Mozilla performance under linux/Xorg

Today I've had enough of some sites being slow under Linux and Firefox. The particular site at this occasion was business.hr. The machine in question is a 3.4 GHz Pentium 4 with a Radeon X300, running the open source "ati" driver (this one loads the actual "radeon" driver). The Linux on the machine is a Debian unstable running xorg core ver. 1.4.1~git20080105-1 and ati ver. 6.7.197-1. I've tried running current git Xorg and drivers, but while I was happy to have it compile, it segfaulted (log file attached bellow).

The site was just awfully, painfully, slide show like slow when scrolling the page. Latest nightly build of Firefox 3 didn't help. My colleague had a ubuntu 7.10 laptop with intel graphics board that worked like a charm on that page. Fortunately we had another laptop with ubuntu 7.04 with intel that exhibited the same problem as my X300 machine. So I decided upgrading it to 7.10 to see if it will go away. I did that and enabled bling (compiz). The problem went away on the newly upgraded laptop.

I figured that enabling compiz on the desktop machine could help. So I've found instructions. Compiz didn't work (gtk-window-decorator didn't draw any window decorations). But the business.hr site scrolling got fast! It turns out that the Option "XAANoOffscreenPixmaps" "true" was the key. I made profiles using sysprof.

The slow one (Firefox 2):


The slow one (Firefox 3):


The fast one:



We spent 81% of the CPU time in X server in the slow case and 40% of the time in the fast case.

I wanted to test another scenario with graphics so I found this site . Even with the naughty XAAOffScreen pixmaps off, it was jerky under Firefox 2, and worked much more smoothly under Windows Firefox 2. So I was delighted to find that Firefox 3 made it smooth under Linux too. Hooray for Linux! Wanting to see if the Firefox 3 was faster still under Windows, I was amused to find that Firefox 3 was jerky there now. Still, under the best case scenario for both systems (ffox 2 on win32, and ffox3 on linux), Windows seemed better CPU usage wise (35% vs 60% CPU usage).


There is clearly more room for optimizations, but it seems quite nice now with Firefox 3 and the XAA stuff fixed. In addition to this quirk, with intel driver on the current Xorg git working nice only when EXA is disabled with Option "AccelMethod" "XAA", Linux distribution vendors certainly must be careful with how they will tune their Xorg configs and patch their drivers.


UPDATE:

Option "AccelMethod" "EXA" also seems to work for the X300, without having to do a Option "XAANoOffscreenPixmaps" "true", which makes sense because XAA and EXA are different acceleration paths. But I'm not sure about the accross the board performance and stability of EXA on the X300. I'll make a new post when I make the tests.

Here's a snippet of the profile with EXA:



UPDATE 2

The addition of the Option "XAANoOffscreenPixmaps" "true" also helps with the intel hardware I have. Be sure not to use EXA on the intel because it also makes things really slow for now. Maybe EXA will be fixed some day, but currently it's better not to use it on intel.

Here are additional links and files.

sysprof:
business-xaaNOoffscreenpixmaps.sysprof.bz2 (fast)
business-xaaoffscreenpixmaps.sysprof.bz2 (slow)
business-ffox3.sysprof.bz2 (slow)

cube-firefox2.0.bz2 (slow)
cube-firefox3.0.bz2 (fast)

business-exaX300.sysprof.bz2 (fast)


Crashing Xorg GIT:
xlog.log

Bug report on the freedesktop bugzilla:
https://bugs.freedesktop.org/show_bug.cgi?id=12069