Slashdot Mirror
Help Beat on Our New Server
Allright the big beefy new VA Research
box (Dual P2/450 with 512 megs of RAM- replacing a dual
P2/266 with 256 megs!) is up and running
at 206.170.14.76. Go
ahead an smash it around a little bit. I need a little
bit of that Slashdot Effect before I try actually running
Slashdot on top of it. If we don't have any major problems,
I'll probably switch everything over later tonight. The
old server was handling over a half a million pages per day.
Hopefully this one will have muscle enough to support us
while I work on the personalized homepage stuff.
Update: 03/08 05:27 by CT : THUD. that didn't take long.
--
"One World, One Web, One Program" - Microsoft Promotional Ad
"Remember when the U.S. had a drug problem, and then we declared a War On Drugs, and now you can't buy drugs anymore?"
I'm certain you must mean -funroll-workstations, right? :)
Brian Fundakowski Feldman
Trying 206.170.14.76...
Connected to 206.170.14.76.
Escape character is '^]'.
Red Hat Linux release 5.2 (Apollo)
Kernel 2.0.36 on an i686
login:
NOOOOOOOOOOOOOO!!
Isn't Beowulf just for specialized Beowulf-enabled apps?
I'm consistently getting the banner ad and the bar with the Slashdot logo and the story icons, then my transfer rate drops off to nothing and I never get anything else, but the comets keep flying past the "N" and Netscape never says the page is done loading. I'm sitting on switched 100Mb Ethernet hooked to a T3, so it's not a lack of bandwidth on my end. I've seen this problem with the main Slashdot server, too, but it's been a long time since I've seen it do it more than once or twice in a row.
run!
1) Go to the 2.2.2 kernel. It really does speed up the SMP a lot and otherwise make a big difference on a dual processor machine. In addition, the dcache means that it will serve static pages VERY fast (no more walking the whole directory tree just to open a web page).
:-).
2) Make sure you're using a good NIC. The typical VAR machine comes with an Intel EtherExpress Pro 100, which qualifies as a good NIC, but make sure they didn't pawn off some surplus junk out of the back room on you
3) Tune your buffers. You have a humongous amount of memory, use some of that as buffers. In the 2.2 kernel with 512mb of memory, you might want to try using as much as 80% as buffer cache... try "echo '2 10 80' >/proc/sys/vm/buffermem". You can change the last number to vary your percentage. The default, using a max of 60% as buffer cache, is somewhat inappropriate for a web server with 512mb of memory, even with a lot of dynamic pages.
I think that the above suggestions will probably improve your performance by anywhere from 20% to 75%, depending upon your particular job load. It should in any event make serving your static front page blazingly fast.
-- Eric
Send mail here if you want to reach me.
Since Chris D. threw a hissy fit:
:-). All that aside, donated machines are often retired show machines, and show machines are often built from whatever parts happen to be lying around because they're not intended to go to real customers. Thus my suggestion about checking what kind of NIC is in the machine, in case it is not the quality EtherExpress Pro 100 that VA usually includes in their machines.
The comment about "surplus junk out of the back room" was a joke. You can tell because it has a little smiley after it (like this:
-- Eric
Send mail here if you want to reach me.
Posted by Kahn:
New server is up and happy from where I'm connected.
traceroute to 206.170.14.76 (206.170.14.76), 30 hops max, 40 byte packets
1 100Mb-OC3 (209.144.48.1) 3 ms 7 ms 2 ms
2 209.144.53.41 (209.144.53.41) 8 ms 6 ms 6 ms
3 sl-gw3-kc-11-0-T3.sprintlink.net (144.224.25.13) 11 ms 11 ms 11 ms
4 sl-bb10-kc-0-2.sprintlink.net (144.232.2.45) 13 ms 12 ms 12 ms
5 sl-bb11-chi-4-0.sprintlink.net (144.232.9.118) 21 ms 24 ms 20 ms
6 sl-bb1-chi-4-0-0.sprintlink.net (144.232.0.134) 20 ms 21 ms 21 ms
7 144.232.8.142 (144.232.8.142) 34 ms 33 ms 31 ms
8 165.87.34.161 (165.87.34.161) 52 ms 31 ms 31 ms
9 165.87.13.61 (165.87.13.61) 143 ms 125 ms 123 ms
10 165.87.13.30 (165.87.13.30) 155 ms 156 ms 143 ms
11 165.87.161.73 (165.87.161.73) 151 ms * 125 ms
12 ded1-fe11-1-0.snfc21.pbi.net (209.232.130.4) 77 ms 78 ms 75 ms
13 209.232.138.214 (209.232.138.214) 77 ms 82 ms 83 ms
14 206.170.14.76 (206.170.14.76) 259 ms 234 ms 249 ms
Posted by nitemare:
take a look at these. Using AIX or linuxppc these things roar www.microux.com
Posted by DocNo:
s t=www.freebsd.org
:)
>some of the webserver survey counters carry stats
>on who is using what OS
Try this one:
http://www.netcraft.com/cgi-bin/Survey/whats?ho
Much more interesting
Do we really need to throw more hardware at this site? I know that the occasional oopsie has brought the site down, but with the last upgrade as well, people have said its the bandwidth stupid whenever things slow down. I'd like to see some stats like load, mem usage etc
Dual P][266 sounds formidable enough. I could see the use for more ram to handle simutaneuous accesses, but that much more CPU?
the new server loads just the top banner ads, and then kinda slows down. Maybe a bit of lighter fluid in the ol' fdd?
"shop smart:shop s-mart" ash
Funny, I wonder what this 4-way Xeon box behind me is running then....
... as you've handily demonstrated.
Linux is a desktop operating system; ask Linus, or any of your other advocates. What you need is a server operating system.
Like this.
You know, it occured to me that this might
be a good time to try to benchmark the
new server, using Linux 2.0, Linux 2.2, and
FreeBSD, just to see how they all compare.
(I tried posting this from work, but it didn't
get through. Then again, I was an AC at the
time).
What do y'all think?
Phil Fraering "Humans. Go Fig." - Rita
(currently testing something about signatures here)
Servlets are good because
:)
- they're to write
- they allow easy accessibility to other applications/machines through Java's enterprise API's such as Enterprise JavaBeans, Java Transaction Service, Messaging Protocol, RMI and JavaIDL (CORBA orb).
- they scale well
- they encapsulate logic well so your web code isn't all gucky with embedded ASP or PHP scripts
- Though if you like the above, Java Server Pages are an option, and actually *are compiled* on the fly to a servlet... so they're more efficient than ASP
Basically I view servlets as an equivilantly good choice to mod_perl. Only thing about Perl is that allows you to make one big mess of code to do something vs. forcing you to cleanly decouple/modularize stuff, which is a big maintenance nightmare over time if you start out with a quick'n'dirty solution (what doesn't ?
out of curiosity, what's better than Mod_perl + apache and a database? ASP+Win NT is "okay", but I wouldn't give it super marks. application servers like Netscape's or IBM's websphere? I think they have a lot of potential myself...
-Stu
Yes it does. Ask Paul. Just because you can't buy it on a CD-ROM doesn't mean it's not running on someone's processor somewhere. No graphics support yet, mind you, but then you were talking about a server, weren't you..?
--
W.A.S.T.E.
W.A.S.T.E.
Oh, bugger it, Paul is at originative.co.uk , not originative.com. Still, he'll probably be thankful to not be spam-botted.
--
W.A.S.T.E.
W.A.S.T.E.
Hey Rob;
Yep, I'm seeing averages of around 500ms from mountain view here at VA off of PA-Cix.
Chris
--
Grant Chair, Linux Int.
VP, SVLUG
Co-Editor, Open Sources
Open Source Program Manager, Google, Inc.
Anyone worth their salt knows that web benchmarks mean kak all - especially when they're designed to differentiate between software. What's important to web software these days is functionality - that's the main reason Apache is more popular than IIS - and staying that way. It's also the reason software like WebSite Pro from O'Reilly can survive in the face of the free IIS.
Matt. Want XML + Apache + Stylesheets? Get AxKit.
[5:57pm] mcope@orion (~): telnet 206.170.14.76 80
Trying 206.170.14.76...
Connected to 206.170.14.76.
Escape character is '^]'.
Connection closed by foreign host.
HTTP is down, other things are up
"L'IT c'est moi!"
...still not finished...groan...I wonder what the bandwidth of the link is?
Good judgement comes from experience, and experience comes from bad judgement.
- W. Wriston, former Citibank CEO
Ack.. www.slashdot.org is slashdotting www.slashdot.org's new slashdot server.
I'd guess it's transfering at 5, 6 bytes per second tops.
But it works.
Shit...
Why RedHat? WHY? WHY!?!?
Damnit, damnit, Sonofabitch.
/me points to the 3rd link from the top on the left hand side of the page, reading "code"
http://slashdot.org/code.shtml
The link is in the top left hand corner of your screen.
faq
hof
CODE
awards
slashNET
older stuff
rob's page
submit story
book reviews
user account
ask slashdot
advertising
supporters
past polls
cachedot
features
topics
about
jobs
BSI
i like i like #)
I've liked Lignux (or LiGNUx or liGNUx) but GLinux is good. We really need a way to avoid the kernel/whole system confusion and both 'Linux' (nonono _DEBIAN_ 2.1, _LINUX_ 2.2!) and GNU/Linux (easy and catchy - not) is bad. Why didn't RMS go for one of the 'GliNUx' or 'LiGNUx'?
I have 100 of these things running....
dt040n05:~# cat beat_up_slashdot.plx
#!/usr/local/bin/perl
use Socket;
require "tcp.pl";
while (1)
{
open_TCP(F, "206.170.14.76", 80);
print F "GET / HTTP/1.0\n\n";
# print $_ while ();
close(F);
}
dt040n05:
I'm certainly doing whatever I can to "help"!
After all, Eddie was designed for doing HA/HP clustering of web servers. Use the bigger iron to handle the MySQL database engine and web serving and then have the old box handle just web serving.
I'd bet that would handle a hell of a lot more than it would by itself....
I am not merely a "consumer" or a "taxpayer". I am a Citizen of the State of Texas
wget -r
works wonders too
Grroooorrr!
...this flame war has gone quite far enough.
--------------------------------------------------------------------------------
----------------------------------------
(Those are supposed to be firewalls.... no I don't know what firewalls really look like unless they're square with wires comin' out the back and that's not the kind I'm talking about.)
Anyway.... let's call truce and stop calling names.
ElpDragon.
yeah actually FreeBSD does support SMP. good job with the facts, moron.
---
---
we stand in life at midnight, we are always on the threshold of a new dawn.
I work for a company that gets over 5 million hits a day on ONE Sun Ultra 2 with 2*200MHz, 1Gb Ram, and 11Gb HDD, with Apache, Oracle, and Netscape serving pages (you have to love corp. politics)
/. gets on a dual 266 is beyond me. Especialy since you are running only apache.
Why you cant handle the load
Oh yeah, before I get flammed for running three webservers on one machine... trust me, if it were up to me, I would run only apache with perl dbi calling oracle on another machine!
Mike
derGott
1 10 ms * 2 3 10 ms 10 ms 10 ms sl-gw10-fw-8-3.sprintlink.net [144.228.134.117]
4 10 ms 10 ms 10 ms 208.12.128.2
5 10 ms 10 ms 10 ms sl-bb11-fw-1-1.sprintlink.net [144.232.11.25]
6 30 ms 40 ms 30 ms sl-bb21-ana-5-0.sprintlink.net [144.232.8.174]
7 40 ms 40 ms 40 ms sl-bb4-ana-4-0-0.sprintlink.net [144.232.1.30]
8 30 ms 40 ms 40 ms lang1sr1-11-0.ca.us.ibm.net [165.87.157.98]
9 30 ms 40 ms 40 ms 165.87.32.114
10 180 ms 180 ms 191 ms sfra1br1-0-0-1.ca.us.ibm.net [165.87.230.202]
11 180 ms 200 ms 221 ms 165.87.13.30
12 201 ms 200 ms 190 ms 165.87.161.73
13 240 ms 251 ms 230 ms ded1-fe12-0-0.snfc21.pbi.net [206.13.28.83]
14 230 ms 240 ms 241 ms 209.232.138.214
15 280 ms 290 ms 321 ms VA [206.170.14.76]
Trace complete.
Tracing route to VA [206.170.14.76]
over a maximum of 30 hops:
1 1 ms 1 ms 1 ms localhost
2 4 ms 2 ms 4 ms router1-ether0.gslink.com [205.157.143.1]
3 9 ms 9 ms 10 ms 209.8.136.110
4 7 ms 7 ms 6 ms fe3-1.mcl1.cais.net [209.8.159.41]
5 7 ms 9 ms 7 ms hssi6-0.me1.cais.net [209.8.159.25]
6 13 ms 11 ms 19 ms mae-east.ibm.net [192.41.177.110]
7 7 ms 9 ms 8 ms beth1sr2-11-0-0.md.us.ibm.net [198.133.27.10]
8 7 ms 9 ms 8 ms 165.87.29.114
9 147 ms 143 ms 126 ms sfra1br2-2-0-1.ca.us.ibm.net [165.87.230.102]
10 127 ms 132 ms 157 ms 165.87.13.42
11 135 ms 140 ms 139 ms 165.87.161.73
12 151 ms 128 ms 127 ms ded1-fe11-1-0.snfc21.pbi.net [209.232.130.4]
13 144 ms 122 ms 145 ms 209.232.138.214
14 * 534 ms * VA [206.170.14.76]
15 * * 598 ms VA [206.170.14.76]
Trace complete.
-- Chad
Do 100 bytes a second constitute the /. effect?
Pinging 206.170.14.76 with 32 bytes of data:
Reply from 206.170.14.76: bytes=32 time=738ms TTL=41
Pinging slashdot.org [206.170.14.75] with 32 bytes of data:
Reply from 206.170.14.75: bytes=32 time=238ms TTL=232
Stan "Myconid" Brinkerhoff
SB.
You Bastards!
"What do you mean, invalid parameters? 9000Gigs of RAM and it can't answer a simple question!" -- Earthworm Jim
I'd be interested. I have a FreeBSD 3.1 box, and a Linux 2.2 box sitting here. For the most part, they peacefully co-exist. (Except for the occasional, "Where the f**k does that config file live in BSD?!?"
I will say that I've noticed a marked improvement in speed going from Linux 2.0 to 2.2, as well as from going from FreeBSD 3.0 to 3.1. (Which is to be expected...3.0 was pretty shaky in places.)
"What do you mean, invalid parameters? 9000Gigs of RAM and it can't answer a simple question!" -- Earthworm Jim
Hmm, that could perhaps be because more time is spent on making FreeBSD actually work properly, instead of making "GNOME GTK++ KDE++ UDE mega-xForms glibc3.1337 RedHat RPM binary ultra-mega-leet" distros.
Connection Timed out over here
Bumm*ED*
I came. I saw. I coded.
yeah southpark!
--Rob
I doubt that I am going to convince a perl hacker to switch to java, but I gotta try:
Have you played around with the java servlet stuff. It rocks. I implemented a "slashdot-like" message board in it, and it only took a couple days. It lacks all the features of slashdot, but it was fast. It was more responsive than static pages, in many cases... (ie all the "nodes" index info was already in memory, so the server didn't have to load anything from disk.) It would probably be even faster with a JIT.
later
--Rob
I am not going to speak for perl, because I don't care for it too much, and don't claim to be any sort of perl guru.
I like java because it is a cleaner language than C++, and it provides features that require some trickery to emulate in C. (ie all that dynamic linking, and the reflection API stuff). Plus it is a very nice language. I really like features like inner-classes. Also gc, objects references rather than pointers, and threads being part of the standard java base classes are a plus.
When I was still in school, I used a lot of C/C++... I implemented a (not-quite-full-featured) JVM in C++, so I knew quite a bit about the underlying JVM, but I hadn't actually used Java to implement anything other than some silly test programs. At my current job, I have implemented are reasonably good sized application (sorry for being vague, but NDA and all...), and I have come to appreciate the java language (and realize how brain damaged C++ is).
I have found that my development time for stuff implemented in java to be lower than equivelent projects implemented in C/C++. Plus, application speed hasn't really been as much of an issue as I had initially anticipated. There definately is a speed penalty, but I think for most things the difference in development time more than makes up for that.
I probably left out some features that I like...
BTW, I used apache + mod_jserv for my little servlet.
--Rob
It's been open source for a long time. Look to your left, click on "code".
----------------------
There is no K5 cabal.
I am not the real rusty.
I guess redhat must compile everything with gcc's -WorkStation flag, "optimizing" it for workstations.
Get a clue.
There's only one version of the apache source, there aren't any redhat specific kernel mods, and they don't include any wacky cron jobs.
--"In dreams begin responsibilities" - Delmore Schwartz
some quotes from my servlet book:
Integration:
Servlets are tightly integrated with the server. This integration allows a servlet to cooperate with the server in ways that a GGI program cannot. For example, a servlet can use the server to translate file paths, perform logging, check authorization, perform MIME type mapping, and, in some cases add users to the server's user database. Server-specific extensions can do much of this, but the process is usually much more complex and error-prone.
Efficiency and Endurance:
Servlet invocation is highly efficient. Once a servlet is loaded, it generally remans in the server's memory as a single object instance. Thereafter, the server invokes the servlet to handle a request using a simple, lightweight method invocation. Unlike CGI, there's no process to spawn or interpreter to invoke, so the servlet can begin handling the request almost immediately. Multiple, concurrent request are handled by separate threads, so servlets are highly scalable.
Also has sections on Power [Premade networking, compresion, database, CORBA, serialization, Third-party components, Serlvet chaining],
Elegence, Extensibility, Flexibility, Portability/Saftey (more targeted at C++)
Rob,
:-(
I think you better check your connections on the server to the 'Net. The response is still a bit on the slow side.
Is the server connected to the 'Net with a T-1 or T-3 line?
Raymond in Mountain View, CA
Doesn't support alpha? I guess thats why there is a little star on the Walnut Creek CD Set that says: "Now with DEC Alpha SMP support, and ELF binaries!"
--
Yes, you CAN slashdot slashdot.
Would that make the new hostname www.slashdot.org.slashdotted.new.slashdot.org?
Or not?
"Although we may build the technology that we define as tools, we must be vigilant that those tools do not define us."
I could be wrong here, but I'm getting basically the same response times from triton, and the 'new' server at .76. Maybe the problem with the load isn't from the lack of hardware, but from a lack of bandwidth? It could also be caused by some code/information bloat on the server, which is simply transfered over to the new server. I don't think Linux is the problem, so I don't need the FreeBSD advocates after me.
:)
My site contains 100% GPL'd source code
mcox.com - Useful Information re: IT, Running, Fitness, Finance, or Ann Arbor!
well, /.'ed the new /. I suppose we should have been a little easier on it on it's first day, but ...
we
It seems like a configuration problem since it loads pretty quickly up until a certain point. c'mon CT!! we're rooting for ya, (no pun intended)
-earl
-- Word of the day: Percussive maintenance is the fine art of whacking the crap out of an electronic device to get it wo
OTOH maybe he ought to give out his root passwd and let the
Just kidding
-- Word of the day: Percussive maintenance is the fine art of whacking the crap out of an electronic device to get it wo
Any thought about open-sourcing it? :)
What, like having a link to the source code as the third item from the top in the left sidebar of every page? Naaah. It'll never happen.
Weblogging Considered Harmful:
So here we are, slowly moving away from the MS monopoly, and then we find OS bigots among us even still? And I thought we were all going to be one big happy family...
Jesus, is it just because people need to point fingers and denigrate others in order to feel superior that we have crap like this happening?
So they're using Red Hat. Big friggin' deal! You want they should use NT? Or MacOS? Maybe they should have got themselves a Sun Enterprise 10000. Would that make you feel better? Yeah, that would be better: skip Linux all together, because they would get shit no matter what distribution they picked. And nobody could complain about a big Sun. That probably would have been the wise choice.
What we need here is a little solidarity. They picked Linux, and that's a Good Thing(TM). All this name calling and devisiveness can only hurt the Linux movement.
Lighten up.
-B
Ash and Hickory, straight-grained and true, make excellent bludgeons, dandy for the cudgeling of vegetarians.
What benchmark software do you recommend? I've punished the living s**t out of a linux box (dual proc, 512MB RAM) with webstone 2.5 and the 100BaseT NIC (EEPro) was the bottleneck for static pages. Funneling half the traffic through a 3c905 (not 905B) boosted throughput quite a bit.
Here are the results from the best run versus Apache (performance goes down with more clients...)
WEBSTONE 2.5b3 results:
Total number of clients: 200
Test time: 2 minutes
Server connection rate: 684.13 connections/sec
Server error rate: 0.00 err/sec
Server thruput: 103.30 Mbit/sec
Little's Load Factor: 198.46
Average response time: 0.290 sec
Error Level: 0.00 %
Average client thruput: 0.52 Mbit/sec
Sum of client response times: 23815.70 sec
Total number of pages read: 82096
Note that these were for the standard Webstone distribution for static file benchmarks. I have no dynamic results. I also don't have a complete benchmark disclosure ready-- this is just a quick test I ran a while back versus linux 2.2.0
Slow as well from here. I'd post a traceroute, but I think it is pretty clear that the problem lies in the server somewhere...
Hmm, slow for me too...about 650ms response vs. 150 for triton, and I'm on a university link.
We get so much "Linux rulezzzzz" and "FreeBSD is for euro-wimpzzz" here on Slashdot. I'd like to see unbiased benchmarks (ha!) comparing Linux 2.0, Linux 2.2, and FreeBSD 3.1.
I've seen some old HTTP server benchmarks comparing Linux 2.0 and FreeBSD 2.8. FreeBSD had twice the transactions per second! Now that the shiny new Linux 2.2 and FreeBSD 3.1 kernels are out, let's get a second look. Does anyone know of any recent benchmarks in this vein??
cpeterso
cpeterso
According to t his Mindcraft Report, NT+IIS running with 256 MB was faster than your Linux+Apache running with 512 MB!
Your Linux 2.2 statistics:
Server connection rate: 684.13 connections/sec
Server thruput: 103.30 Mbit/sec
Mindcraft's NT4 statistics:
Server connection rate: 929 connectsion/sec @ 50 clients
Server thruput: 137.7 Mbit/sec @ 50 clients
Wanker: Were the Linux statistics you posted from your EEPro + 3c905 benchmark? The NT benchmark used two 100Base-TX network interfaces on the server. This might explain why the NT numbers look so good..?
One caveat: the NT benchmark was done with WebStone 2.0.1 and the Linux benchmark was done with WebStone 2.5b3.
cpeterso
I tried several times in a row and could not
connect at all.