Letters to the editor

We couldn't come up with any concrete advice, as Pencom works exclusively with high tech engineers. We think the Web is an easy sell and it always surprises us how reluctant many people/companies are. We recently began a new Web development division and, after building class libraries and finishing our book on Java, will be out soliciting business.

The Web is still new and fresh and it may be a blessing that you can't get work with any of the big companies in this field. There are many people out there starting things up themselves.

Performance Q&A questions and answers

Editor's note:See Adrian Cockcroft's frequently asked questions for more answers to reader questions.

Adrian:

I have incorporated perflog script from your book, Sun Performance and Tuning. It works well but I can't find how to interpret the generic labels sd0, sd1 etc. What can I use to cross-reference to real disk slices?

--Kenneth Woods, US Navy

It is a lot easier to use the SE toolkit though. Read this months column for info on how to download it. xiostat.se prints this:

% xiostat.se
                                 extended disk statistics
disk      r/s  w/s   Kr/s   Kw/s wait actv  svc_t  %w  %b
c0t0d0    5.0  3.0   12.0   19.0  0.0  0.1    9.5   0   6
c0t2d0    0.0  0.0    0.0    0.0  0.0  0.0    0.0   0   0
c0t3d0    0.0  0.0    0.0    0.0  0.0  0.0    0.0   0   0
fd0       0.0  0.0    0.0    0.0  0.0  0.0    0.0   0   0

Adrian:

I have been digging through the IO statistics collected by sar on a database server, and am confused by the numbers. In the Sun performance class, I was told that the IO rate on a system can be determined by the summation of the Block IO rate (buffer cache), Raw IO rate, and Paging IOs. That would be the data items named - bread/s, bwrit/s, pread/s, pwrite/s, pgin/s and pgout/s.

On this system, I did this and came up with an average IO rate of 4.26 IOs per second, which can be broken down as 1.56 read/s, and 2.70 writes/s.

The problem is when I compare this data to the disk activity, collected for each disk. This system has 46 disks, and when I sum the r+w/s values for each time interval, and average them, I get values of 39 IO/s for the same time interval. What is the reason for the discrepancy in these numbers. I can only think that I am missing something.

This system is a Sybase database server (Solaris 2.4), and is using raw partitions for the database. The DB partitions are mirrored by Sybase internally. My suspicion for the discrepancy is that the values reported in the pread/s and pwrite/s field are incorrect, because when I graph this, I found that there are no raw IOs during the business day, and only some activity in the evening. As the DB was in constant use during the day, this must not be reporting the disk raw IOS, but only the tape raw IOs. Is this because the raw devices Sybase is using is not using the physio() interface?? And if so, how can I see the activity?

As an aside, the man page for sar describes the disk IO as being reported in reads/s and writes/s, but what actually is produced is r+w/s.

Thanks for any help you can provide,

--Chris, (firm indeterminate)

I think that Sybase is using async I/O, via a kernel fast-path that does not go via pread and pwrite. I don't think sar has any way of counting aio activity.

Adrian:

Is there anyway to tie wio to a process. Occaisionally I have had problems with excessive wio times. Most of the time the problem has been due to a process that has hung either with a modem or tape drive. Finding the problem was done by elimination. I don't have an immediate problem but would like to be able determine the processes waiting for I/O in the future.

--(name and firm indeterminate)

The OS does not distinguish between waiting for network input or waiting for disk/tape input. It can provide counters for most other kinds of waiting (wait for paging, wait for CPU to become free etc).

Load up the SE toolkit and look at msacct.se, there is no builtin Solaris command that displays the data. The SPARCworks Analyzer is the "official" consumer of this information. Its available as a special ioctl on /proc (PIOCUSAGE) that msacct.se uses.

Sysadmin columnist `surfing with the sharks'

Hal:

I enjoyed your article on Surfing With The Sharks and have discovered much information in my quest to raise awareness within the company I work for.

We have a difficult situation, management wise, in regards to information security in that it is often seen as counter-productive and treated accordingly.

It is a real challenge for me to continue on in the hopes that I can improve their image of security professionals. I would be interested in any other similar articles or research material you know of that you are willing to pass on.

--(name and firm indeterminate)

There was a LISA conference a few years back on management and dealing with non-technical managers, I think it was LISA VII. call the usenix office in Berkeley (or check its home page, www.usenix.org) for more details -- you can order back issues of the proceedings.

Dan Appelman's tutorial is also quite good -- if you make it to LISA X in Chicago this year, I think he's giving it again.

RAID questions continue

Brian:

We have a SPARC storage array on a machine running Solaris 2.4, however there are no drivers. Are the drivers available from an FTP site, or must they be purchased?

--(name and firm indeterminate)

If you have Solaris 2.4 11/94, you should have gotten drivers on a CD with your storage array. If you have any later version of Solaris 2.4, the drivers are on the the Solaris CD itself. (Eg, Solaris 2.4 HW 3/95.)

Either way, be sure to get the latest patches.

Brian:

Thank you very much for your article. It answered some key questions at exactly the right point for myself and my colleagues. A co-worker was a week away from spending $26K uneducated dollars after a cute dog and pony show by a couple of HP salesmen. Your article helped illustrate that RAID would not address at least 75% of the result of our redundancy v. access time equation.

--(name and firm indeterminate)

Brian:

Your article was informative and useful. Toward the end of the discussion on RAID-5, I was wondering, If I wanted to remove a failed disk and restore operation, leaving my site continuing to operate, I saw no clear way that this could be done. Disk mirroring seemed to provide a solution to this problem.

Can you offer thoughts on hot restoration of service using RAID-5?

--(name and firm indeterminate)

RAID-5 is fully capable of this type of operation. When a disk fails, the RAID-5 volume operates in degraded mode, providing data generated via the parity function. The failed disk can then be replaced and resync'ed, also using the parity function.

The big differences between mirroring and RAID-5 in this context are:

• substantial performance degradation for RAID-5 degraded mode
• complexity of reconstruction (mirror = just read it, RAID-5 = read lots of stuff and XOR it).
• cost of redundant configuration

Brian:

I would like to know if there is a tool to analyze the performance of a RAID system in order to improve the response time. this tool see the RAID like a single disk? this tool can be used for the all different RAID system (0-5)? this tool run under unix SVR4 ? this tool suggest values for the different tuning parameters (chunk..)? etc...

--(name and firm indeterminate)

1. There isn't a single one available across all platforms. If you have AIX you can run an I/O trace against the affected device, although that takes considerable fiddling with the trace output. Solaris 2.5 has a internal system trace facility too (known as tnf, see the prex man page). This also is a bit difficult, but we have successfully used it to discover very detailed performance information. I think that both of them are quite a bit more complex and less easy to use than what you are after.

2. The whole idea of a RAID volume or metadisk is to rpesent the appearance of a single disk but with different performance and/or reliability characteristics. Any tool that could measure this would act in this way.

3. "Could such a tool can be used for the all different RAID systems (0-5)?"

   No such thing is available today. The RAID advisory board (RAB) is developing a RAID benchmark, intended to be sort of a "rhaidstone" but it is not generally available.

4. The RAB benchmark suite only reports performance, it does not suggest tuning. HP's AutoRAID claims to do this but it does not really implement very much of use to real users. One of the most difficult problems here is that device usage almost always varies widely, and what tunes best at one moment doesn't usually tune best all of the time.

From our reader support dept.

Well, I was interested if there even is a comparison at this point. Meaning is the Pentium-pro chip even starting to break into this market of the SPARC workstation? Is this maybe an alternative to the high prices of a SPARC Workstation? I have yet to hear a reply from anyone that may have an unbiased answer. This question is specifically in regards to the Intel Pentium-Pro 200 running Windows NT Vs a SPARC 10 Workstation running Solaris 2.5. Now I am pretty sure that the SPARC would be faster, but how much faster is my question?

--Jim Martel, (firm indeterminate)

Well, what's faster: Honda or Toyota?

Back when we editors worked on a magazine called Advanced Systems we had a test lab that would conduct these types of comparisons all of the time. We were the first to compare the first PowerPCs from IBM vs Pentium.

Intel's benchmark results, which are tweaked, unverified SPEC92s, suggest the Pentium Pro bests almost all RISC chips in integer performance, and is average in FP. But there's more to life than integer and floating point. And there's much more to life than Intel-supplied benchmark data.

In our Advanced Systems days, we ran many benchmarks, including several that put a strain on I/O. There are also graphics benchmarks, which a computer powered by an UltraSPARC CPU should shine on.

There are relatively few users, really, that need THE FASTEST computer at a given price point. Those who do know who they are. For the rest of us, we pick the apps we need, then see which computers appear to run them best at the price we can afford.

Editors:

We have connected our SPARC-5 machine to the Ethernet network. On the same network we have some 50 PCs connected. We also have another PC based Unix machine connected which is acting as access provider for all PCs. Through this PC based Unix Server (mail-relaying machine), we are able to send/receive files through a dish-antenna to CyberSpace. Each PC user has an unique ID. The SPARC-5 also has an unique ID. The SPARC-5 is configured as a Server and it has sub-users who used to share the same machine.

The SPARC-5 users are able to send mail by telneting to the PC-Based Unix server and then by using the mail command.

My question is, how do we send/receive mails/files directly (and not by telneting) to the Cyberspace? Do we have to setup any mail-related .cf files?

Please give me some solutions.

--Manix, (firm indeterminate)

Our guess is you need to configure sendmail on your mail relay machine, and perhaps on the SPARCstation-5 server as well. I recommend you read a basic Unix system administration book, and buy a copy of the Sendmail book from O'Reilly and Associates. The O'Reilly's URL is http://www.ora.com

Good luck. Configuring mail properly is a difficult task.

Editors:

I have download a few FAQs about Solaris, but still don't have an answer to my simple question: How do I make a TCP/IP connection? Which is the step-by-step procedure to get it right?? I have the Solaris x86 desktop box, an a HP NetServer LC (pentium, 16Mb, 1GB, CDD, 3.5d) I don't use NIS or NIS+. I trying to use the files of the host.

Why the changes I make under openwin with admintool, not seem to work???

--(name and firm indeterminate)

Your question is a difficult one to answer, since there is much you've left out of your message. It's akin to asking us "Why doesn't my fax machine work?" Where does one begin to answer such a question?

We recommend you go to a book store and pick up a copy of Hal Stern's TCP/IP book (published by O'Reilly & Associates).

We run admintool as root.

Editors:

This is a suggestion for a feature in $500 Network Computer (NC) being promoted by Sun. I think this feature would increase the usefulness of such a computer many folds.

I am excited by the possibility of NC (Network Computers) for $500. One of the reasons that I think would excite a lot of people like me would be able run some type of remote access for MS-Windows comparable to "Laplink" and "PcAnywhere" on NC.

I use this remote access technology routinely on my $2,500 Pentium PC to connect me to my office PC over the modem which then gives me access to my office computer system worth several grands.

I would think remote access like that of "Laplink" and "PcAnywhere" is something that does not require a lot of computing power and it certainly does not require any local storage. This feature alone in an NC would increase its usefulness so much that many people would embrace it to "bring their office home for $500."

What do you think of NCs providing this feature?

--(name and firm indeterminate)

What do we think? Right idea, wrong implementation. The Internet Toaster is a network device. Without a net it's a useful as a netless X terminal. Barring any high-speed networking revolution this summer, the first Toasters will be used on LANs. As higher-speed networking connections reach into homes, perhaps we'll see Internet Toasters as consumer devices.

Telecommuters will probably be the first home Internet Toaster users. They won't need Laplink or PcAnywhere because the files manipulated remotely will be the files on the server -- the same files accessed at work.

SysAdmin: How to synchronize workstation time

Security: Secure your Solaris Web server

Performance Q&A: What is swap space?

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

URL: http://www.sunworld.com/swol-04-1996/swol-04-letters.html
Last updated: 1 April 1996

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

URL: http://www.sunworld.com/swol-04-1996/swol-04-letters.html
Last modified: