Readers Speak Out: Letters to the Editor This month: A critique of an HTTP server survey mentioned in our Webmaster column, plenty of performance questions for Adrian Cockcroft (with answers from Adrian), and more

July  1997

Mail this article to a friend

Webmaster

To Chuck Musciano:

Regarding your recent Webmaster column, "Who's really winning the browser and Web server wars?" (http://www.sun.com/sunworldonline/swol-05-1997/swol-05-webmaster.html), we found that the Netcraft survey is bogus to the extent that it counts each Web page root as a server.

Many of the servers listed as Microsoft are actually just one server with virtual connections. One of them is a site in the U.K. that listed about 200 colleges and schools. It all runs on one box, therefore it's really one Web server. It gets 200 listings for Microsoft.

We also made Microsoft give us listings of sites they knew had IIS3. None of them were secure sites, none of them were commerce, and of the two banks they gave us one was a page pointing to the bank that bought them, and the other was static HTML only. All of the other pages were on Apache or Netscape. I think the bank had gotten pressured into trying it.

We also found that most MS sites are static sites only -- no real techie stuff happening.

John Allen

Chuck replies:

Multi-hosted servers are not unique to Microsoft-based sites, and thus will inflate the counts for all servers that support multi-hosting. I don't think this invalidates the Netcraft survey results, but readers should be aware that multihosting can skew the results somewhat.

As to your contention that nothing "techie" is going on within many Microsoft sites, that has little bearing on the raw count of servers in use. Even static sites deliver valuable content. Your urge to bash Bill is laudable, but let's make sure that we sling barbs that are sure to stick.

Career Advisor

To Edgar Saadi:

On reading your Career column in the June issue: "Sysadmins: Sharpen your people skills!" (http://www.sun.com/sunworldonline/swol-06-1997/swol-06-career.html) it seems to me that not one of your system administrators really answered or addressed the letter by "Frustrated."

The fact is that s/he does not need to "straighten out" his or her boss as much as s/he needs to talk to him or her. Frustrated should be telling the boss what s/he told you:

"our users do drive me up a wall: They don't read their manuals; they constantly interrupt me; and they act like their problem is a big crisis."

I bet the boss would try and help. Here are a couple of things that I say to appease my users:

"I am sorry I cannot help you right now because I am working on a [insert problem description]. E-mail me with the details of your problem, and I will get back to you as soon as I can. If you really need help, contact the lead SA, and s/he will try to help you."

"Sorry, I'm pretty busy right now -- but if you look at the man pages for the command x you should find your answer. If not, send me an e-mail, or call me, and I will see if I can take a minute or so to help you out."

These statements tend to put the problem back in the user's court without making them upset that you are ignoring them. Always remember that to users their problems are a high priority, so you need to explain why you cannot help them. And you must reply to e-mails in a prompt manner, or at least pass them on to another SA or the lead SA to follow up on.

Good luck to Frustrated; I think we've all been there at one point or another.

Tim Feeney

Readers respond to Igor

First off, it's a bit strange that your "Letters to the Editor" page, does not have an e-mail URL to send one! [Editor's note: Thanks for pointing this out. As you can see, we've now added our e-mail address to both the top and bottom of our Letters page.]

Now to visit the specific reason that came up: I am replying to the letter (http://www.sun.com/sunworldonline/swol-06-1997/swol-06-letters.html) from Igor, asking for help on PPP and Solaris.

You should have pointed out to him that since he recently BOUGHT Sun PPP, he gets free 30 day (or longer?) support. Sun doesn't advertise this much; surprise, surprise.

Philip Brown

Another response for Igor

While some may find it heretical, an excellent primer on PPP and Linux/Unix can be found in the Linux PPP-HOWTO pages. While some of the finer details may differ between Solaris and Linux, the general concepts will be same for both operating systems. Try http://www.redhat.com/linux-info/ldp/HOWTO/PPP-HOWTO.html to begin with.

Good luck,
Wally Wharton

Performance Column: Q&A with Adrian Cockcroft

To Adrian Cockcroft:

I would like to lower the minfree on some of our large drives. Is there any way to see this as a block count? I know that tunefs will change it, but I would like to see the particular filesystem's params.

Dru Nelson

Adrian replies:

You can calculate from the numbers in df -k by looking at the percentage used, the size and the in-use totals.

CPU phenomenon

To Adrian Cockcroft:

Thanks for your marvelous Sun Performance and Tuning book and SunWorld column. I have a performance phenomenon that I can't explain: I use percol to monitor a busy SMTP Gateway and it reports that CPU sometimes reaches Amber and Red. But at the same time, the CPU usr/sys are only five and 28 percent respectively.

My interpretation is that the CPU is idle while the run queue is still long. Is my interpretation correct? If so, what is happening? The SMTP Gateway is a E3000 with four 167-MHz CPU and one-gigabyte memory, serving about 60,000 mail accounts.

Thanks,
Andrew

P.S. -- I'm waiting for the updated SP&T.

Adrian replies:

Your CPU usage is bursty; during bursts the queue is long enough to slow things down, but the gaps in between reduce your overall average.

For SMTP delivery, this is unlikely to be an issue, as you are unlikely to be concerned about user response time as long as mail gets delivered through the system without too big a delay.

Source code minefield

To Adrian Cockcroft:

Do you know where I can obtain the Virtual File System information for Solaris? Not white paper stuff but the actual implementation detail. If so, does it contain any examples? If not, can I purchase the source code for one?

We want to create a virtual file system on Solaris 2.5.

Thanks in advance,
Robert Fradette

Adrian replies:

This is a real minefield. The interface is subject to change in each release, and SunSoft is very reluctant to freeze and document it for third parties. A few companies have managed to do it, but it's harder than it should be.

Sun now pricelists Veritas VxFS, which is the main third party version I know of, but I think Veritas paid for source code access.

What's your special need?

Workaround: Ultra servers with Multipacks

To Adrian Cockcroft:

It seems that Ultra servers no longer support NVSIMM. Do you have any recommendation for users of Ultra servers with plain disk systems (e.g. Multipack) in a I/O intensive environment?

unsigned

Adrian replies:

Dedicate a disk as a logging drive, make a 100-MB partition at the beginning of the disk and don't use the rest of the disk for anything (this is to insure that the heads never seek outside the first 100 MB). Use ODS4 to set up a transaction log for your striped filesystems (metatrans).

I recently did this on some 4.2-gigabyte, 7200-rpm drives, and the sequential speed of the drive is high. It worked very well, you just have to ignore the "wasted" 4.1 gigabytes of space. After all, you're using all the performance of the drive, and they're fairly cheap and non-volatile.

The other alternative is to use NVRAM in the disk controller (SSA or RAID units), but with Multipacks this is the best you can do.

Disk 10 performance

To Adrian Cockcroft:

I am running into performance problems with an E3000 functioning as a WWW proxy. I'm trying to handle a moderate load (1.5 million hits per day) and have had problems with Disk IO performance. I am running 600 proxy instances, and under moderate loads the system spends most of its time in iowait state, and the service times (as seen from iostat) go sky high (~700ms).

Is there anything I can do to optimize the system for this type of task, or, am I hitting the performance limits of the Ultra's SCSI IO system?

Thanks,
Todd Ignasiak

Adrian replies:

You don't say how your disks are set up, but you need to either use a storage array or other NVRAM based RAID unit to do fast writes or Disk Suite to set up an extra disk as a transaction log (metatrans). You'll need to use only the first 100 MB of the disk to insure that the head never seeks outside the first few tracks.

If you don't have a spare disk, pick the most idle one you can find, and use that in the meantime. Stealing 50 MB from the swap space on the system disk into an extra partition to use as a log for your cache stripe is one way to do it.

Sun memory allocation

To Adrian Cockcroft:

Can you help me?

We have a program that processes a very large binary file (711 MB). To help to speed up the processing of this file our programmers added a user configurable cache that sits between the main processing loop and the file system. Our machine has half a gig' of memory and when the program is running the code, data, and internal cache probably consumes 300 MB. The program then runs for a very long time (multiple days); the file is opened at the beginning and not closed until the job is complete.

While this processing is going on, our available memory count goes down, which I guess is due to the filesystem cache, we then get to the point where free memory hits around the 5000 mark (output from vmstat). Then it starts to page, user cpu usage drops, system CPU usage hits around 10 percent, and it's idling around 88 percent. At this point we have loads of free swap space.

I have a feeling that eventually our program (it's memory cache) gets into contention with the file system cache, and it starts the thrash. Other Unix boxes we have tried this on don't seem to suffer like this, then again, they don't claim to have the same file system cache mechanism - one in particular being the Digital Unix offering.

Any help/advice that you can provide would be great.

Martin Waller

Adrian replies:

My May 1997 column, "The memory go round," (http://www.sun.com/sunworldonline/swol-05-1997/swol-05-perf.html) should help. In the meantime, increase lotsfree so that the reclaim cuts in earlier and softer.

Since the job runs for days, does it keep rereading the big file? If not then you may just have a process that is too big for memory. Try reducing the configurable cache size to a minimum.

Beyond Unix?

To Adrian Cockcroft:

In the May 27, 1997 edition of PC Magazine there's an article entitled "Beyond Unix" (p.167) where they review a number of Web servers running on various platforms including Solaris 2.5.1.

In their tests, Netscape Enterprise and FastTrack servers running on Solaris performed very poorly in comparison to the equivalent Windows NT version of the product. The difference in performance between the Solaris and NT hosted Web servers was striking.

In light of the great articles you've written about tuning Solaris and Web servers, would you care to comment on the strange results reported in PC Magazine?

Thanks!
Nigel Simpson,
Chief Architect,
InfoMation Publishing Corp.

Adrian replies:

There is an "official response," but the basic problem is that we optimized for SPECweb96, which is an open agreed benchmark, and we've tuned Solaris 2.6 to run it well. The PC Magazine review used Solaris 2.5.1, which has been out for ages and is known to be untuned.

They also used a non-standard benchmark that was not able to drive the Sun properly to full speed.

To remedy the situation, and prevent it from happening again, we now have internal PC-oriented benchmarking facilities, so we can run and tune for the PC tests. We are also close to releasing Solaris 2.6, which is a big step up and outdates any Web performance measures using earlier releases.

Solaris 2.5.1 has been fast enough for most users and is also very stable. www.sun.com was up continuously from August 1996 to Jan 1997 without a reboot; after a hardware upgrade it's been up continuously since.

Solaris 2.6: Linear scale, multiple CPUs

To Adrian Cockcroft:

I am benchmarking a Web server on a PC server with four Pentiums at 133 MHz. But the number of CPUs doesn't affect the performance of the Web server when three CPUs or more are used.

Somebody suggested that it is due to network saturation.

My testbed is 10-Mbit Ethernet and connected building-wide. It's not possible to have a separated network here. Maximum throughput produced by WebStone is about 4.15 Mbit/sec, and it's on a two CPU configuration.

How can I be sure the network isn't saturated?

According to an article, perfmeter in Open Windows cannot provide precise information about network condition. And as far as I know, packet field on perfmeter just provides relative load information. Is this true?

Any comment or answers will be welcomed! Thank you.

Tae-sung Kim
Operating Systems Lab.
Dept. of Computer Science & Engineering
Korea University
Seoul, Korea

Adrian replies:

Your network could be saturated, but the Solaris 2.5.1 Web server does not scale beyond two CPUs. You need Solaris 2.6, which ships next month and scales linearly.

To the Editor

To the editor:

I would just like to express the need for this site. In the war to keep open standards (and Java is one of our last hopes), central information databases are needed. SunWorld is one of the best I have found. The world with Microsoft in more control of the market would be a very gray place. Let open standards win!

Scott Ahlf

If you have technical problems with this magazine, contact webmaster@sunworld.com

URL: http://www.sunworld.com/swol-07-1997/swol-07-letters.html
Last modified: