This is part of The Pile, a partial archive of some open source mailing lists and newsgroups.
Subject: Re: Best Hardware for Linux System?
From: rslomkow@parts-unknown.com
Date: Thu, 02 Dec 1999 20:22:54 -0800
Jackie Sladky wrote:
> We have a 'small' machine that we use as a development system. We have
> Linux 5.2, JDK 1.1.7, Oracle 8, Apache Jserv 1.0 and Redhat now.
>
> We want to get a powerful machine for production web pages. We've heard
> that we can get either a two CPU intel machine or a one CPU Compaq Alpha
> machine. We are not sure which one would be better for Linux, and at
> sometimes in the future (6 months or one year), we will need to expand the
> system.
As much as I like to support chipsets other than Intel, for systems
that you expect to run commercial software packages on I would say GO
INTEL. Intel has much wider support of binary packages both
pre-compiled and commercial to choose from. As for opensource
software, it doesn't matter, and alpha does have some advantages of a
clean 64 bit implementation.
Consider purchasing a VA Research or Penguin Computing server, for a
server platform that will give you knowlegable linux support.
I would also reccomend a Linux distribution that includes glibc 2.1 and
a 2.2 kernel.
===
Subject: Re: Best Hardware for Linux System?
From: Rick Moen <rick@linuxmafia.com>
Date: Fri, 3 Dec 1999 11:21:12 -0800
Quoting rslomkow@parts-unknown.com (rslomkow@parts-unknown.com):
> As much as I like to support chipsets other than Intel, for systems
> that you expect to run commercial software packages on I would say GO
> INTEL. Intel has much wider support of binary packages both
> pre-compiled and commercial to choose from. As for opensource
> software, it doesn't matter, and alpha does have some advantages of a
> clean 64 bit implementation.
Of course, one intriguing option, in cases where you're running multiple
machines on a LAN, is to have _one_ local x86 Linux box, and then being
free to run other architectures for all the other boxes.
You see, people often forget that the X Window System is a networked
protocol. That makes x86's "wider support of binary packages" a
non-problem:
Let's say you've bought and installed Bentley Systems's Microstation
package (a CAD program) for x86 Linux. However, your LAN also sports
a Macintosh G4 running LinuxPPC, a Netwinder, a DEC Alpha running Red
Hat 5.0, an Amiga running Debian, and a DECstation 500 running NetBSD --
and you sometimes like to work from those. No problem: ssh into the
x86 box (with ssh's X forwarding enabled), and launch Microstation.
It images onto your local machine's X server.
For that matter, if one doesn't happen to _need_ a lot of x86-only
binary programs, an Alpha might be just the ticket. Certainly, given
the choice Jackie described (2-CPU x86 vs. 1-CPU Alpha), the Alpha's
horsepower is likely to be much more generally accessible. (SMP is
no match for a high-end single-CPU system.)
===
Subject: Re: Best Hardware for Linux System?
From: "Jeffrey B. Siegal" <jbs@quiotix.com>
Date: Fri, 03 Dec 1999 16:17:18 -0800
Rick Moen wrote:
> No problem: ssh into the
> x86 box (with ssh's X forwarding enabled), and launch Microstation.
> It images onto your local machine's X server.
If you have a 100mbit (or faster) network, and you are concerned about
performance more than passive wiretapping of the X data on your local network,
it is better to forgo encryption and just set DISPLAY on the target back to
your own X server's address.
> Certainly, given
> the choice Jackie described (2-CPU x86 vs. 1-CPU Alpha), the Alpha's
> horsepower is likely to be much more generally accessible. (SMP is
> no match for a high-end single-CPU system.)
True, but depending on the application, a *one* of dual
Intel processors may well be as fast or faster than a single
Alpha. Of course, it depends on the particular CPU models
in question, and there are some types of computation where
Alpha really shines, but Intel has largely caught up to
Alpha in recently. If your application is single-thread
compute intensive, you may also want to consider an Athlon
(near-Alpha performance with Intel binary compatability).
===
Subject: Re: Best Hardware for Linux System?
From: Rick Moen <rick@linuxmafia.com>
Date: Fri, 3 Dec 1999 16:32:30 -0800
Quoting Jeffrey B. Siegal (jbs@quiotix.com):
> If you have a 100mbit (or faster) network, and you are concerned about
> performance more than passive wiretapping of the X data on your local network,
> it is better to forgo encryption and just set DISPLAY on the target back to
> your own X server's address.
<shrug>
Try it both ways. I use ssh more or less reflexively, at this point.
One element fewer to worry about. There are tradeoffs in this picture,
obviously: ssh can (through compression) save bandwidth, at a small
cost in CPU overhead. On the other hand, you could use LBX to cut
bandwidth.
> True, but depending on the application, a *one* of dual Intel
> processors may well be as fast or faster than a single Alpha.
In my experience, this tends not to be the case for equivalent-cost
machines, but that of course depends on what kind of deal you get
for each.
I'm eagerly awaiting ramp-up of Athlons, so I can reasonably compare,
but have not yet had a chance.
===
Subject: Re: Best Hardware for Linux System?
From: "Jeffrey B. Siegal" <jbs@quiotix.com>
Date: Sat, 04 Dec 1999 10:02:54 +0000
Rick Moen wrote:
> Try it both ways.
I have been involved with extensive tests of X performance
over various network transports, including ssh, with and
without LBX. While not noticable for simple apps, ssh
simply can not come close to keeping up with 100 mb
bandwidth. This makes a major difference for complex apps
that transfer a lot of data (particularly images). With
compression, the picture gets *worse* (due to additional CPU
requirement), not better, particularly if either CPU is
doing anything else CPU-intensive.
> ssh can (through compression) save bandwidth, at a small
> cost in CPU overhead. On the other hand, you could use LBX to cut
> bandwidth.
Both LBX and ssh compression are not advised for a fast LAN,
particularly one with a decent switch (making using less
than the full available bandwidth usless). But if your
network bandwidth is scarce, you should try them. In this
case, the best combination is usually to enable ssh
compression and disable zlib compression (but retain
X-specific compression) in lbx with the -nocomp option.
YMMV.
===
Subject: Re: Best Hardware for Linux System?
From: "Jeffrey B. Siegal" <jbs@quiotix.com>
Date: Sat, 04 Dec 1999 10:07:39 +0000
Rick Moen wrote:
> In my experience, this tends not to be the case for equivalent-cost
> machines, but that of course depends on what kind of deal you get
> for each.
Well, I was looking for something about six months ago to do
solve large linear programming problems. Initially I
dismissed Intel for this workload, but based on feedback
from svlug, I did some benchmarks and that performance was
*better* on both moderately high end P-III and very high end
Intel (Xeon) than equivalent-cost Alpha. I found that to be
a win because with Intel binary compatiliby, why screw
around with anything else? Maybe the situation is different
today. Haven't looked at the new offerings closely.
===
Subject: Re: Best Hardware for Linux System?
From: Rick Moen <rick@linuxmafia.com>
Date: Mon, 6 Dec 1999 08:31:41 -0800
Quoting Jeffrey B. Siegal (jbs@quiotix.com):
> Well, I was looking for something about six months ago to do solve large
> linear programming problems.
Which would have been particularly sensitive to math co-procoessor
performance, and therefore skew the comparison.
===
Subject: Re: Best Hardware for Linux System?
From: Rick Moen <rick@linuxmafia.com>
Date: Mon, 6 Dec 1999 08:35:49 -0800
Quoting Jeffrey B. Siegal (jbs@quiotix.com):
> I have been involved with extensive tests of X performance over
> various network transports, including ssh, with and without LBX.
> While not noticable for simple apps, ssh simply can not come close to
> keeping up with 100 mb bandwidth.
Because your hosts become processor-bound? How odd! I almost
never see processor-bound Unix boxen. (Actually, remove the "almost"
from that.) You must be really hitting that math chip.
===
Subject: Re: Best Hardware for Linux System?
From: "Jeffrey B. Siegal" <jbs@quiotix.com>
Date: Mon, 06 Dec 1999 14:06:00 -0800
Rick Moen wrote:
> Because your hosts become processor-bound?
No, because current processors (even otherwise idle) can not
do encryption and compression (ssh -c blowfish *without*
compression comes close) at 100 mb/sec, and even if they
could, the process still adds substantial latency, can make
a big difference to X performance.
> You must be really hitting that math chip.
If you're referring to my other post about linear
programming, no. The X performance tests were done on
completely different systems.
===
Subject: Re: Best Hardware for Linux System?
From: "Jeffrey B. Siegal" <jbs@quiotix.com>
Date: Mon, 06 Dec 1999 14:06:42 -0800
Rick Moen wrote:
> Which would have been particularly sensitive to math co-procoessor
> performance, and therefore skew the comparison.
Right, but that used to be the area where Alpha's had the
biggest advantage, until Intel (largely) caught up....
===
Subject: Re: Best Hardware for Linux System?
From: Rick Moen <rick@linuxmafia.com>
Date: Mon, 6 Dec 1999 17:48:19 -0800
Quoting Jeffrey B. Siegal (jbs@quiotix.com):
> Rick Moen wrote:
>> Which would have been particularly sensitive to math co-procoessor
>> performance, and therefore skew the comparison.
>
> Right, but that used to be the area where Alpha's had the biggest advantage,
> until Intel (largely) caught up....
Pardon me if I'm a little incredulous. FPU has been _Intel's_ strength,
for a long time. That's what that company is so sore at AMD over the
Athlon -- it's hitting them where they live.
===
Subject: Re: Best Hardware for Linux System?
From: Rick Moen <rick@linuxmafia.com>
Date: Mon, 6 Dec 1999 17:51:12 -0800
Quoting Jeffrey B. Siegal (jbs@quiotix.com):
> Rick Moen wrote:
>> Because your hosts become processor-bound?
>
> No, because current processors (even otherwise idle) can not do
> encryption and compression (ssh -c blowfish *without* compression
> comes close) at 100 mb/sec, and even if they could, the process still
> adds substantial latency, can make a big difference to X performance.
In other words, processor-bound. OK, obviously your mileage differs.
I've never seen a *ix box in ordinary operation be in that state.
It is unclear what variety of latency you're referring to in this
context, but never mind.
===
Subject: Re: Best Hardware for Linux System?
From: Javilk <javilk@polly.mall-net.com>
Date: Mon, 6 Dec 1999 18:46:13 -0800 (PST)
> In other words, processor-bound. OK, obviously your mileage differs.
> I've never seen a *ix box in ordinary operation be in that state.
I am not sure I understand how you are measuring this, or what,
exactly, you mean. (Not to argue with you, just to clarify what we are
talking about. I would like to learn more about this metric.)
Although my boxes stand idle most of the time, once a week they spend
half a day with the CPU utilization at between 75% and 99.9%
according to TOP. Does that qualify as being processor bound by your
definition?
Or are you implying that cache runs dry, and the CPU waits at a lower
level than TOP can discern? That I believe!
I will agree with you that most boxes spend most of their time
waiting. That is sort of inevitable unless you have oodles of cache,
oodles of RAM and a really fast swap drive which is separate from the rest
of the disk space. However, many ISPs seem to run tight enough that the
lousy response time (seconds per character,) imply high CPU utilization.
(Though it could also be high I/O channel utilization, particularly on
swap.)
===
Subject: Re: Best Hardware for Linux System?
From: Dan Lyke <danlyke@flutterby.com>
Date: Mon, 6 Dec 1999 20:49:54 -0800 (PST)
On Mon, 6 Dec 1999, Rick Moen wrote:
> Quoting Jeffrey B. Siegal (jbs@quiotix.com):
> > Right, but that used to be the area where Alpha's had the biggest
advantage,
> > until Intel (largely) caught up....
>
> Pardon me if I'm a little incredulous. FPU has been _Intel's_ strength,
> for a long time. That's what that company is so sore at AMD over the
> Athlon -- it's hitting them where they live.
In my experience the Alpha has long been the best floating point per $ and
per MHz of anything around. My day gig for the past 4.5 years has been at
Pixar, which knows a little bit about CPU utilization and floating point.
And "a long time" means < 7 years, because back in the early part of this
decade Weitek and Cyrix were handily whupping Intel at the floating point
game on Intel's own turf, even after Intel integrated the floating point
into the main processor with the 486DX there were add-on processors to do
math faster that got used with AutoCAD and similar apps.
I'd argue that FPU has *never* been Intel's strength. They've made a set
of middling chips that did nothing excellently, but did *everything*
tolerably well, and they weren't stupid enough to latch on to the RISC
bandwagon. That's why they've won, while others were busy trying to fit
into this niche or that, Intel's been striding down the middle cutting as
wide a swath as possible.
===
Subject: Re: Best Hardware for Linux =?iso-8859-1?q?System=3F?=
From: Markus Gutschke <markus@gutschke.com>
Date: Mon, 06 Dec 1999 23:07:43 -800
> Pardon me if I'm a little incredulous. FPU has been _Intel's_ strength,
> for a long time. That's what that company is so sore at AMD over the
> Athlon -- it's hitting them where they live.
I always heard a somewhat different story, although I cannot find
the relevant numbers to verify which story is true.
Basically, I have always been told, that Intel traditionally had
crappy FPU performance in comparision to most typical workstation
processors. This started to become a problem right around when Intel
was developing the original Pentium processor, because this deficiency
kept them from entering the workstation market.
So, for the release of the Pentium, Intel invested significant
resources to improve FPU performance considerably. This left all of
their competitors in the PC market in the dust, and brought Intel's
CPU's into the same ballpark region as the other workstation vendors.
Still, the DEC Alpha (now Compaq) was considered a superior processor
for all FPU related problems. It also had the added benefits of being
available at higher clockrates than pretty much anything else, and of
supporting 64bit architectures.
The latter is less of an issue with Linux, because I have heard that
Linux on Alpha cannot really make use of the entire address space and
currently iA32 is actually a better choice if you need large amounts
of memory.
The other advantages of the Alpha are also starting to become less and
less significant.
Another traditional advantage of workstation platforms over PCs is
their improved bandwidth to various subsystems (memory, I/O,
graphics,...). While I believe this is still somewhat true, PC
architectures benefit from turning a lot of high-performance
technologies into commodities and they thus succeed in (almost)
catching up with traditional workstations, but at a considerably lower
price point.
So, if you need absolute raw performance and money doesn't matter,
go ahead and get quotes from Compaq, Sun, IBM, HP, ... and find
the workstation that fits your exact needs. Otherwise, save the
money by buying a PC; make sure that you invest in quality parts.
And if you have money left, buy a second machine and distribute
the load between the two. YMMV
===
Subject: Re: Best Hardware for Linux System?
From: Rick Moen <rick@linuxmafia.com>
Date: Tue, 7 Dec 1999 11:56:47 -0800
Quoting Dan Lyke (danlyke@flutterby.com):
> I'd argue that FPU has *never* been Intel's strength. They've made a
> set of middling chips that did nothing excellently, but did
> *everything* tolerably well, and they weren't stupid enough to latch
> on to the RISC bandwagon. That's why they've won, while others were
> busy trying to fit into this niche or that, Intel's been striding down
> the middle cutting as wide a swath as possible.
Well, in any event, long digressions about ssh and FPUs aside, I recall
saying that, depending on what sort of pricing one finds -- and on
whether one has a _real_ need for x86-compatibility in that _one_ box --
an Alpha-based system might indeed give superior performance for the
dollar. I would certainly consider one.
===