[ILUG] NT vs. Linux comments from linux-kernel

[John P. Looney]

 Some comments I thought were informative from linux-kernel:

-----------------------------------------------------------

From: Ingo Molnar <mingo at chiara.csoma.elte.hu>
Subject: Re: Dual Ethernet Slowness Causing Lousy Benchmarks?
On Tue, 29 Jun 1999, Jordan Mendelson wrote:

> It seems to me that all the benchmarks used dual ethernet controllers
> when they were benched against NT, and that's when the major degredation
> showed up in performance (at least that's what this C'T article
> suggests).

If by 'major degradation' you mean 'it can only deliver 1500 hits/sec',
then yes. I mean, it must suck to be an OS that can serve only a meakly
129 million web pages per day on a single box ;)

Seriously, there is absolutely no 'major degradation' [i have the newest
c't issue in front of me], it's just that we do not scale too well
compared to 1-CPU performance (only 30-40% improvement with 4 CPUs).

Such nonscaling is absolutely unacceptable to have around - so David S.
Miller has SMP-scaled the networking code in 2.3 already - on my box and
2.3.9 TCP bandwith scales almost linearly on localhost - it shows almost
no scaling on 2.2.10. (i do not have 4 100mbit ethernet connections - yet)

------------------------------------------------------------

From: Jordan Mendelson <jordy at wserv.com>                                                       
Subject: Re: Dual Ethernet Slowness Causing Lousy Benchmarks? 

Ingo Molnar wrote:
> Such nonscaling is absolutely unacceptable to have around - so David S.
> Miller has SMP-scaled the networking code in 2.3 already - on my box and
> 2.3.9 TCP bandwith scales almost linearly on localhost - it shows almost
> no scaling on 2.2.10. (i do not have 4 100mbit ethernet connections -
> yet)

This is what I see from the C'T article (approximates):
Linux 1 CPU 4K files: 950 requests/second
NT    1 CPU 4K files: 940 requests/second

Linux 1 CPU 8K files: 850 requests/second
NT    1 CPU 8K files: 710 requests/second

Linux 1 CPU CGI: 100 requests/second
Linux 4 CPU CGI: 240 requests/second
NT    1 CPU CGI: 25 requests/second
NT    4 CPU CGI: 52 requests/second

Now here's the kicker:
Linux 4 CPU Dynamic Data/2 Eth: 1500 requests/second
Linux 1 CPU Dyanmic Data/2 Eth: 1100 requests/second
NT    4 CPU Dynamic Data/2 Eth: 2600 requests/second
NT    1 CPU Dyanmic Data/2 Eth: 2100 requests/second

 Now scalability in SMP isn't the problem here, as even with 1 CPU we can't
beat NT's 1 CPU. It is either Apache's handling of the dynamic data
generated by the CGI, NT's version of perl is simply faster than ours,
Linux's scheduler hurting the perl process needed to delier dynamic data,
or somewhere, the network load balancing is not working and we are maxing
out on bandwidth or CPU. 

 Now I'd say it's the CGI, but there was another CGI test which didn't
suffer like this, although it didn't generate the same dynamic data from
what I can tell. 

These numbers also show that NT's SMP is about as bad as ours is. Is this a
problem with Intel's SMP itself or what?

From: Marc Mutz <Marc at Mutz.com>
Subject: Re: Dual Ethernet Slowness Causing Lousy Benchmarks?

Jordan Mendelson wrote:
>
> Now here's the kicker:
>
> Linux 4 CPU Dynamic Data/2 Eth: 1500 requests/second
> Linux 1 CPU Dyanmic Data/2 Eth: 1100 requests/second
> NT 4 CPU Dynamic Data/2 Eth: 2600 requests/second
> NT 1 CPU Dyanmic Data/2 Eth: 2100 requests/second
>
s/Dynamic Data/single 4K file/g;
Marc

------------------------------------------------------------------

From: Ingo Molnar <mingo at chiara.csoma.elte.hu>
Subject: Re: Dual Ethernet Slowness Causing Lousy Benchmarks?

> Now scalability in SMP isn't the problem here, as even with 1 CPU we
> can't beat NT's 1 CPU. It is either Apache's handling of the dynamic
> data generated by the CGI, NT's version of perl is simply faster than
> ours, Linux's scheduler hurting the perl process needed to delier
> dynamic data, or somewhere, the network load balancing is not working
> and we are maxing out on bandwidth or CPU.

thats a misunderstanding. It's _not_ CGI scripts - thats nonsense anyway,
we are over 1000 hits/sec.

Yes, IIS serves static files faster than Apache ATM - although i'd like to
see the e-commerce site that serves one single 4k file over 4 100mbit
ethernets ;) And even in the static-serving tests, Linux beats NT hands
down if the number of files is increased to a more realistic value.

I am getting 1-CPU performance of 3500 hits/sec (small static files) with
a 'dumbed down' webserver-demo coded up by Zach Brown which uses
sendfile() and Linux's async IO API. Apache 2.0 (judging from Dean
Gaudet's plans) will be able to get such numbers as well.

------------------------------------------------------------------

 Seems that NT can saturate 4x100mbit network cards. Linux/Apache can't,
due to poor SMP scalibility for network I/O (which is allegedly fixed), and
the fact that the test was measured using the same 4K file, repeatedly.

 Yet. New linux-specific versions of Apache can get 250% current
performances with static pages (this is using sendfile(), not khttpd).
Network code in 2.3 has gotten rid of all outstanding scalibility issues.
All other measurements, Linux whipped NT's ass.

 Seems as usual, Linux wins if you are willing to use the "Not just yet,
but real soon now" response...

Kate

-- 
Microsoft - the best reason in the world to drink beer