Letters to the editor -- SunWorld, October 1995">
Letters to the editor
I got a copy of your Performance and Tuning book last week, and from what I've had a chance to read so far, it looks great. Well-written, straight-forward, and very useful. Thanks!
(Our new Sun 1000 PDB system is arriving tomorrow)
--John Sellens, U of Waterloo
I was browsing www.sun.com looking for information on the SPARCArray (I know you care but I'm just being conversational) and I noticed Sun has a link to your performance tool on its main page.
I downloaded the ruletool packages:
and installed them. They were MOST simple to install and they appear to work without modification (which is more than I can say about some of the software I have BOUGHT from Sun). Many thanks!!! I have the graphic ruletool running on my screen as I write this.
va_monitor (and ruletool) are complaining about two things on my system and I was wondering how to change the fs? or kernel? settings to the recommended values. I have a slow disk (too many system partitions on an old Micropolis). I will take care of the slow disk, but I was wondering what to do to implement the suggestion to bump DNLC entries from 1127 to 2254.
Adrian detected slow disk(s): Tue Sep 12 16:35:48 1995
Move load from busy disks to idle disks State disk r/s w/s Kr/s Kw/s wait actv svc_t %w %b delay amber c2t6d0 7.8 3.5 38.9 30.9 0.2 0.3 42.7 3 17 481.2Adrian detected Directory Name Cache problem (amber): Tue Sep 12 18:00:22 1995
Poor DNLC hitrate, increase ncsize DNLC hitrate 85.9%, reference rate 82.42/s DNLC has 1127 entries, try increasing it (and inode cache) to 2254
--Mark Holcomb, Driftmier Engineering Center
You need to add
set ncsize=2254 set ufs_ninode=2254
to /etc/system and reboot.
In fact, your hit rate isn't all that bad, and the performance benefit is small (thats why this only ever gets to amber, not red state) so you could ignore it. If your system is an NFS server you should increase it though.
While looking at your description of ruletool and the rules it uses, I was very pleased to see you point out the basic problems associated with the CPU stat called 'IO-Wait time'. I work for Cray Research in their Business Systems Division. Basically we design and manufacture a high-end SPARC server that is targetted at a customer when the largest SC2000 is not large enough. We run Solaris 2.4 and are fully compatible with the sun4d architecture. I have worked quite a bit with George Cameron and other kernel/performance engineers at SMCC.
As a software engineer in our operating systems group, I must have explained this misleading 'IO-Wait time' statistic hundreds of times to our analysts, techincal support group and customers. It is such a frequent issue that I have a mail template with the 'IO-Wait time' lecture in it.
My questions is: Could Sun be persuaded to drop this kernel stat OR at least have it report zero on a multi-processor system? Vmstat already combines 'IO-Wait time' and 'regular idle time' in its display. I would find it more accurate to say that we only have User, System and Idle time on our SC2000. IO-Wait time isn't really a CPU state on these machines, so why pretend it is?
I welcome you're ideas on either side of this issue. Maybe you feel that it does have some merit. I personally find it more misleading than informative, but I am open to debate.
I have found your books and articles very useful in my work and appreciate your time.
--Robin Williamson, Cray Research
Unfortunately Unix, with all its crufty detritus gathered over the years, is too entrenched to make it easy to remove or change measures. I'm talking to several groups at Sun about what measures to add in future releases, and would welcome your input.
Hats off to a GREAT column. After a magnificent book, only a column like yours could top it off! I would like to see more details of ALL tunable KERNEL parameters with a description of each parameter -- like a man page. You did a similar job on page 196 of your "Sun Performance and Tuning" book. I would like to see a description of sar output applied to SPARCs. Some of these info is on your SPT book but its scattered, how about a sar book ? Perhaps an updated manual page describing not only its output, but threshold applied to SPARCs.
On you book very little is mention about TCP/IP tunning. Is this because Solaris has little or no tunable parameters?
I am a Unix contractor and my customer has five (soon to be 12) SPARC 2000s
running Sybase. I'll be more than happy to supply
info of "real" data for your column or book.
--Ed Silva, (firm indeterminate)
I'm thinking of basing a Q&A column on the "what are the kernel tunables for Solaris" question.
The reason there is little on TCP/IP is that I'm not a networking expert. There are plenty of networking experts out there, and I've specialized in looking for areas where it seems there is a lack of expertise.
Very recently, I've become heavily involved in WWW server performance issues, and I'm learning more about TCP/IP tuning. I have to restrict the time I spend on networking and database issues, since either of them could consume all my time if I tried to get up to speed.
In your SunWorld Online article, "Tune That Code," you suggest preferring vfork(2) over fork(2) if the only thing the child does is to exec a new image.
However, in the Solaris 2.4 Reference Manual, the documentation for vfork(2) states: "This function will be eliminated in a future release. The memory sharing semantics of vfork() can be obtained through other mechanisms."
A word on those "other mechanisms?"
("See my article in ..." would be a more than acceptable answer, of course. :)
Other than that, thanks for a very educational and useful article!
--Yuval Kfir, Indigo Ltd., Israel
mmap is the other mechanism the docs refer to.
I've recently read your article "system performance monitoring" in SunWorld Online, and downloaded the SE toolkit. Ruletool is very nice, but I have one problem. It seems that when response times are slow, all states are in the green (or white) but the CPU is only 0-15% idle. I know the rules for CPU performance are based on the run queue length, and that RUNQ_BUSY defaults to 3.0. How is this number established, and how can I adjust it to properly reflect my system performance?
Also, do you know of any tools/methods of monitoring network traffic
more closely (if at all possible)? For example, separating LAN traffic
from WAN traffic, show transmission rates, etc.
--BJ, (firm indeterminate)
A good tool for monitoring networked server systems, and breaking out the network traffic, is AIM Sharpshooter. http://www.aim.com
I've downloaded your SE toolkit recently. IT IS A GREAT TOOL!!!! Finally someone from Sun decided to take care about area which is very important to users but COMPLETELY neglected by Sun. In my opinion (and it is my company's official view) this tool HAS TO BE INCLUDED in future Solaris releases! (I don't care if I have to pay extra for this tool!).
Please forward this letter to your "upper levels"!
--Dragan Jurkovic, Les Aliments Dainty Foods
I'm surprised you overlooked us as a player in this industry. Maybe that's our fault, but nonetheless, we have been around for 13 years and support many on-line services with our engine, including AOL, Prodigy, Dow Jones News Retrieval, and Data Times IQ. In 6/94 we introduced our Web tool and are support many more sites than Verity, Dataware, Architext and Conquest combined. Check out some of our flagship Web sites:
Thanks for your timely article on CDE. We are in the middle of evaluating Triteal's CDE offering for our HP systems. Until I tried it, I was rather impatient to see CDE in Solaris.
My initial impression of CDE is that it doesn't live up to the hype. Basically HP VUE on steroids, it seems to be as difficult to administer on more than a few machines as VUE is. Customization is easy but limited. Users still have to edit resources files for things not covered by the Style manager!
In the provided applications, it seems that CDE retains the "yuck" and loses some of the nice features of the applications it inherits. The calendar manager is still the same one-host-based paradigm which causes so many problems with permissions. The text editor is very stripped-down, without the oh-so-useful 'text extras menu' (run the selected text through various filters) that some of us rely on (this concept doesn't appear anywhere in the interface). The mailtool, while having more functionality than Sun's (it can speak MIME as well as Sun attachments), leaves much to be desired in the UI department (why can't I type in a mailbox name without using menus?) and in the fancy features (I like to be able to edit my mail on the fly; and I don't like the way the detached-message-window works in CDE). I won't go into the file manager beyond the lack of flexibility in the information included when I want to include date, size, etc.
The question is, what is the chance of any of this changing? My guess is that there is none. So I'm glad Sun is not going to force it on us.
--Andrew Tefft, GE
I also want to commend you on the quality of the articles in your September issue. I especially enjoyed the articles on CORBA and the new C++ draft standard.
I look forward to your next issue.
--Steve Totten, Software Systems Specialists
Your (Debunking meths) article summarizes very valuable information; in fact, it is worth keeping this article around to sensitize junior software engineers or managers to the importance of using a sound methodology with or without tools.
The time is ripe for someone to write a comparative review of software design tools & methologies backed where each toolset and methodology is backed up by at least one reputable team using it. That way, it would allow readers to decide which team they are closest to or aspire to emulate and feel confident that following that team's choices worked for them.
--Name indeterminate, NASA
One question: When talking about HP and Distributed Smalltalk, you mention that C++ bindings have just been approved. From context, I think you might mean Smalltalk bindings. Also, I think C++ bindings have been approved. Or maybe you meant C++ bindings have just been approved, and I am mistaken (i.e., only C bindings were previously approved).
--Tom Ioerger, University of Illinois
I've just finished your lead story. Much appreciation to those who thought of the readers' need for immediate gratification: hyperlinks were judiciously sprinkled in all the right places.
Regarding Netscape's grudging compliance with U.S. export laws -- how about a follow-up story on encryption standards and how U.S. munitions laws are shooting itself in the foot?
--Gerald Gold,(firm indeterminate)
In the hardcopy magazine, I could quickly glance at the new products and pick out keywords about products to determine if I had any interest. With the format in the online version, I have to actually call up the link to determine those keywords.
Please consider including a few summary words about the product so that I can decide up-front whether or not I want to follow the link.
--(Name and firm indeterminate)
The type size it too small for comfortable reading, and I'm not familiar enough with HTML to know whether/if it can be readily changed.
I like the feedback request at the end of each article, but I don't think it's an entirely accurate mechanism. Of the few articles I've read already, they've all been excellent and just the right length, but unless I go to read something that doesn't interest me, you won't get feedback from me on that article and if I read something that doesn't interest me, then it will get a lower rating than it perhaps deserves. You might want to rethink the wording of the inputs or put an explanation of what kind of feedback you are expecting from your readers.
It's a great idea, though! Don't get rid of it.
Reiterate the rational, systematic, basics for the old hands and move the newcomers up to speed. Thanks. The second article was a little sparse, but the bibliography was a fine touch.
If you have problems with this magazine, contact email@example.com
Last updated: 1 October 1995
If you have technical problems with this magazine, contact firstname.lastname@example.org