COPYRIGHT (C) 1984-2005 MERRILL CONSULTANTS DALLAS TEXAS USA
Newsletter FORTY
/* COPYRIGHT (C) 1984-2005 MERRILL CONSULTANTS, DALLAS, TEXAS */
****************NEWSLETTER=FORTY****************************************
MXG NEWSLETTER NUMBER FORTY dated Feb 14, 2002.
Technical Newsletter for Users of MXG : Merrill's Expanded Guide to CPE
TABLE OF CONTENTS Page
I. MXG Software Version 19.19.
II. MXG Technical Notes
III. MVS Technical Notes
IV. DB2 Technical Notes.
V. IMS Technical Notes.
VI. SAS Technical Notes.
VII. CICS Technical Notes.
VIII. Windows NT Technical Notes.
IX. Incompatibilities and Installation of MXG 19.19.
See member CHANGES and member INSTALL.
X. Online Documentation of MXG Software.
See member DOCUMENT.
XI. Changes Log
Alphabetical list of important changes
Highlights of Changes - See Member CHANGES.
COPYRIGHT (C) 2002 MERRILL CONSULTANTS DALLAS TEXAS USA
I. MXG Software Version 19.19 is now available.
1. Major enhancements added in MXG 19.19:
See CHANGES.
II. MXG Technical Notes
1. Benchmarks of Data Transfer and MXG Processing on PCs.
Early results were presented to the MXG User Group Meeting at CMG
2001 in Anaheim, December 5, 2002.
a. Measurement of sustainable data transfer rates between platforms
and between disk drives, with the theoretical capacity of two LANS
is given in this table:
Table of Sustainable Data Transfer Rates
KB=KiloBytes MB=MegaBytes GB=GigaBytes=1,073,741,824 bytes
Connection/path KB/sec MB/sec MB/min MB/hr GB/day
Dial In 56kbit 6 .006 .4 20* 1
T1 1.44mbit 144 .144 8 500* 12
MVS to 10 mbit LAN 470* .50 30 1800 42
10 mbit LAN @100% 1024 1.0 60 3600 85
750MHz 100mbit LAN 2.4* 144 8640 202
500MHz C: to D: IDE 6.5* 390 23400 550
100 mbit LAN @100% 10240 10.0 600 36000 850
750MHz C: to D: IDE 12.0* 720 43200 1000
850MHz C: to D: SCSI 24.0* 1440 86400 2000
Those measures are for raw data transfer bytes.
Note that a T1 connection transfers 500 MegaBytes per hour, while
a 100 mbit LAN transfers over 8500 MegaBytes per hour (using about
25% of the theoretical capacity). The much higher disk-to-disk
transfer rates on the PCs show that data transfer time, and not
the MXG run time, will be the constraint on daily processing time.
b. Download SMF data, Execute BUILDPDB on three different PCs:
The SMF file contained only SMF 6, 26, and 30 records:
842 MegaBytes, Compressed to 72 MegaBytes, 11 to 1.
On a 2064/1C5 CPU (SU_SEC=10540) BUILDPDB took 18.5 minutes
elapsed (and 6 minutes CPU).
To download SMF data, you must first convert it from RECFM=VBS to
RECFM=U (copy with IEBGENER), and then (optionally) zip (compress)
the file on MVS (to reduce the download time), and then unzip on the
PC and process the uncompressed SMF file. The transfer was over a
T1 line and the timings in minutes for the processing are these:
Phase 500 MHz 2 SCSI 850 MHz 2 SCSI 1.3MHz 1 IDE
No-ZIP Zipped No-Zip Zipped No-Zip Zipped
RECFM=U 1 1 1 1 1 1
ZIP on MVS 9 9 9
ftp download 89 7 89 7 89 7
unzip on PC 1 1 2
BUILDPDB 19 19 10 10 38 38
total minutes: 109 37 100 28 128 57
With 11:1 compression to reduce the transfer time, the 850 MHz box
with two fast SCSIs would take 28 minutes to process (including the
download), as compared with 18.5 minutes for BUILDPDB on MVS.
Note that the speed of the PC and its disk drives has significant
impact on the total run time.
Zipping on MVS: "Info Zip MVS", SHAREWARE:
Get program/documentation at ftp://ftp.info-zip.org/pub/infozip/mvs
Input limit: One full Volume uncompressed; input must be a member
of a PDS (RECFM=U conversion outputs to PDS member).
Unzipping on PC: "Zip Magic for Windows", $40 USD purchase:
Get program/documentation at (the address has changed several
times, as the product has changed hands) as of Feb, 2003:
http://www.aladdinsys.com/win/zipmagic, $39.99, downloadable.
When installed on PC, can read the zip file directly with MXG, so
there is no need to unzip (and hence you only need 72MB of disk
space for the downloaded zipped SMF file, instead of an additional
842 MB for the expanded SMF file!). This is really slick!
Mar 30, 2004 Note: ZipMagic is NOT supported for Windows/XP,
although it has worked without error on many files. If it does
fail with XP, you have no recourse with the vendor, who has not
responded as to their plans (if any?) for XP support. Their
alternative Stufit product doesn't support direct read of zips.
MXG Member DOCZIP (in MXG 19.08+) contains the JCL and specifics for
the installation of both "InfoZip MVS" and "Zip Magic for Windows".
c. Detail examination of impact of Compression options.
The TYPE30 program was used to compare compression costs, since it
reads the SMF file and writes out the SAS datasets in a single SAS
data step. The input dataset could be zipped or unzipped, and the
output SAS datasets could be written compressed or uncompressed.
MACHINE: 1.3 GHZ ONE IDE DISK:
Input Output Elapsed mm:ss User CPU System CPU Total CPU
RAW NO 13:22 2:22 :33 2:55
RAW YES 11:27 4:30 1:45 6:15
ZIP NO 11:00 2:19 :54 3:13
ZIP YES 10:23 4:30 2:04 6:34
Comment: Fastest ET was with both zipped input and compressed out.
Compressing Output saved 2 minutes.
Compressing Input saved 2.5 minutes.
Compressing both saved 3.0 minutes, or 22% savings.
Here, the elapsed cost to do I/O is greater than the
elapsed cost of compression, and CPU was available.
Best Time: 10:23.
MACHINE: 500 MHZ TWO SCSI DISKS:
Input Output Elapsed mm:ss User CPU System CPU Total CPU
RAW NO 5:12 4:57 :11 5:08
RAW YES 5:35 5:15 :13 5:28
ZIP NO 5:52 4:49 :55 5:44
ZIP YES 6:09 5:12 :52 6:04
Comment: Fastest ET was with unzipped input and uncompressed out!
Run time is more than halved due to faster I/O.
Here, the processor is CPU bound, and adding compression
to a CPU-bound process increases the run time.
Savings from slowest to fastest was 57 seconds, 30%.
Best Time: 5:12.
MACHINE: 850 MHZ SAME TWO SCSI DISKS:
Input Output Elapsed mm:ss User CPU System CPU Total CPU
RAW NO 3:00 2:47 :08 2:55
RAW YES 3:12 3:00 :09 3:09
ZIP NO 3:27 2:48 :34 3:22
ZIP YES 3:36 2:59 :34 3:33
Comment: Faster CPU with same disks; run time dropped from 5 min
to 3 minutes, but processor is still CPU bound, so fastest
run time is still unzipped and uncompressed.
Savings from slowest to fastest was 36 seconds, 16%.
Best Time: 3:00.
MACHINE: 850 MHZ ONLY ONE SCSI DISKS USED FOR INPUT AND OUTPUT:
Input Output Elapsed mm:ss User CPU System CPU Total CPU
RAW NO 3:48 2:49 :09 2:58
RAW YES 3:38 3:00 :09 3:09
ZIP NO 3:59 2:49 :34 3:23
ZIP YES 3:45 3:00 :34 3:34
Comment: One disk is slower than two disks; contention increased
run time from 3 minutes to 3:38, but CPU was the same
(and these numbers show the consistency of CPU measures).
Savings from slowest to fastest was 36 seconds, 16%.
Best Time: 3:38, with output compressed (probably due
to Write cost being much higher than read cost).
The overall conclusion is that zipping and compression is always good
when you have CPU available; even when compression does increase the
run time, the increase is small, and more than justified by the large
saving in disk space and transfer time.
But faster CPUs and faster Disk drives have a bigger impact on run
time than does compression!
2. Benchmarks comparing SMF and Web Log processing on Linux & Windows:
==========================================
SMF (842MB unzipped type 6, 26, and 30s.
119MB compressed output, 54MB Sorted)
Elapsed Elapsed Data Copy Sort
mm:ss sec sec sec sec
Dell Inspiron 1.3GHz Win2K 12:44 764 738 18 08
IBM ThinkPad 1.3GHz Linux 6:26 386 347 15 24
IBM ThinkPad 1.3GHz Win2K 3:06 186 152 18 16
Desktop Clone 866MHz WinXP 3:28 208 183 7 16
Comments:
Both IBM and Dell are 1.3 MHz CPUs; 2:1 difference might be due to
disk drives, but I think is actually due to better floating point
execution on IBM; see below.
Linux to Win2K, 2:1 difference shows Win2K NTFS outperforms the
nfs that comes with Red Hat Linux 7.2.
In all cases, the data step is much longer than the sort time; this
has always been the case with MXG processing.
==========================================
WebLog (93MB unzipped input, 672MB uncompressed output dataset,
95MB compressed, 484343 observations 1456 obs length,
SORTSIZE=256M, only one variable in BY List runs:)
Elapsed Elapsed Data Copy Sort
mm:ss sec sec sec sec
Dell Inspiron 1.3GHz Win2K 4:08 244 55 13 180
IBM ThinkPad 1.3GHz Linux 8:25 505 93 28 384
IBM ThinkPad 1.3GHz Win2K 3:43 223 55 15 163
Desktop Clone 866MHz WinXP 4:07 247 73 12 162
In all cases with this character data, sort times are much longer
than the data step time to read the input and create the dataset!
I had never seen this behavior, because SMF data is highly numeric.
This caused me to examine what caused the longer sort times, and it
was the length of the BY list that was first discovered to have a
major impact on the sort times.
850 866 850 866
MHz MHz MHz MHz
DATA STEP SORT STEP
secs secs secs secs
All Variables in BY list: 75 73 612 914
Only one var in By List: 75 73 176 162
And those were for the best runtimes with compression enabled; the
uncompressed 850MHz sort time with all variables was 831 seconds,
(and copying compressed took 11 seconds, copying uncompressed 204).
But fortunately, since we really need to be able to sort on more
than one variable, my webmaster found that SYNCSORT for WINDOWS had
been released, and it appears to be a solution if you are going to
sort large character datasets under SAS on Windows:
All Variables, SYNCSORT: 75 73 90 --
and clearly, SYNCSORT under Windows outperforms the internal SAS
sort provided in SAS Version 8 under ASCII, with this kind of
highly compressible character data.
On unix, SYNCSORT is available, and SAS does interface with some
unix systems, but Linux is not yet one of them; SAS Technical Note
TS-6830 documents the systems that support SYNCSORT as host sort.
For SAS Version 9, SAS expects to support SYNCSORT under Linux,
HP/UX, Solaris, and Tru64. And when SYNCSORT supports the 64 bits
with AIX 5.1, SAS intends to support that as well in Version 9.
SAS Version 9's internal sort will be a portable multi-threaded
parallel sort that is much faster on SMP hardware than the current
internal SAS sort, so improvement is planned there.
On MVS, SAS sites have either SYNCSORT or DFSORT installed, so I've
never had reason to measure its internal SAS SORT performance.
Comments:
1. Note that Dell and IBM are much closer with all character data
whereas IBM was much faster with SMF numeric data.
III. MVS Technical Notes.
20. Al Sherkow's posting to MXG-L answers many questions about how to
and how not to set up WLM policy:
On Friday, January 11 Joan Kontra asked about manipulating CPU
resource among LPARs. One option Joan was considering was to
manage their workload by via changes in WLM. Part of my response
included:
"I do not recommend changing WLM policy. Develop a policy and
stick with it. It may need to evolve as your workloads change over
time, but switching policies is (in my opinion) not a strategic
decision."
Backchannel I received some questions about my email and setting
up WLM policies:
1. How can you have different levels of service as a function of
shift?
2. Can they use the WLM to change service between her daytime and
nighttime LPARs?
3. What's wrong with having the operators change the policy at
those times of day? Isn't that also just a console command?
While WLM was always supposed to manage your workload within your
sysplex in my opinion WLM is only now, with z900 & zOS, beginning
to do that. Most of WLM's control was within an image/LPAR. The
exception being the transaction "spraying" capabilities that could
choose which of a group of images to send a unit of work to.
Part of the problem WLM faces is that once a transaction/batch
job/query begins in an image it stays on that image until it is
complete. It doesn't matter if higher priority work arrives, and
it doesn't matter if high priority work is already running on the
system. In general, work does not move. Now consider another LPAR
in the sysplex that has available resources. It would be nice if
work could move from busy LPARs to less busy LPARs.
With z900 and IRD if the LPARs are on the same CEC then IRD will
move the CPU resources and/or I/O resources from an LPAR not using
them to an LPAR that can use them. Remember z900, z/OS and IRD
introduced the new term "LPAR Cluster". All the LPARs on the same
CEC in the same sysplex are an LPAR Cluster.
Today's service policies take advantage of how WLM has been
working. Consider CICS and Batch. In many policies CICS has a
higher importance than batch. Generally production is in one LPAR
and development/test is in another. Still everything is fine,
because CICS is higher than batch. Now z900, z/OS and IRD come
along and these partitions run in a single CEC. The production
CICS and production batch are running fine, meeting their
objectives and goals. Development CICS starts missing its
objectives. The WLM policy has always been operating at a sysplex
level, but IRD provides new capabilities that truly extend the
policy across the images within the LPAR Clusters. With IRD, the
policy is now implemented at the Sysplex level .... so that
development CICS has the same importance as Production CICS and is
higher than all batch (including production). IRD takes resources
away from the production LPAR to help the development LPAR's CICS.
Viola! Finally WLM is working throughout the sysplex.
I recommend that sites need to determine the requirements and
service levels of all their workloads at all times. Also remember
that service is service, regardless of timeframe. I think you need
to perhaps consider your "timeframes" when defining the policy and
use some other means to get work into the correct timeframe
buckets. In this case "The workload on 'System A' consisting of
one LPAR is always to be favored during the day". I wonder if she
means the whole LPAR or something it is running? Their other
requirement "The workload on 'System B' consisting of 2 LPARs is
to be favored at night sometimes."
Perhaps the 'System A' work during the day, and the 'System B'
sometimes favored could be at the same importance. One critical
question that has to be answered: "What would they do if these
workloads were running at the same time?"
On Changing Policies:
Even before WLM I was generally opposed to changing IPS at
different times of the day. In much of performance and managing a
data center there are alternative techniques to solve a problem
that will lead to the same conclusion. I think that analyzing data
from a site that changes SRM parameters (IPS, ICS, OPT or WLM) is
more difficult. There can be the added problem of an offshift IPS,
and prime shift IPS and dealing with last night's outage. While
catching up do you run in 'offshift' or the 'prime' configuration?
For this reason I prefer a comprehensive 'all time frame' view of
an installation's service.
This is also greatly effected by the available capacity. Rarely is
an upgrade in capacity the size that the site actually needs. The
engines of today's machines are just too big. I'm not sure anyone
runs out of MIPS in 250MIPS chunks. Today if I'm upgrading a z900
the steps are about 250 MIPS each. What works well after an
upgrade may not work well just before the next upgrade. This is
the time that parameters are "tweaked" to solve specific problems,
often without regard to overall service. In IPS terms this is when
people tried to use time slicing or played with the time slicing
pattern. (Should we favor a 40% or the pattern or 50% of the
pattern?) After the upgrade few installations go back and
"untweak". So over time the "tweaks" build up. I've often found
this when called in to work on performance problems. Peter Enrico
has done papers at Share, z/OS Expo and CMG about periodically
reviewing your WLM Service Policy to see if it is still doing what
you want. This addresses the same problem.
19. IBM APAR PQ55355 for Websphere Application Server V4.0.1 for z/OS
and OS/390 lists many problems in type 120 SMF records that are
in error and corrected. PQ55355 is associated with SERVICE LEVEL
W401009 of WebSphere Application Server V4.0.1.
18. BMC MAINVIEW CMF originally from Boole and Babbage type 73 values
for PCHANBY and PNCHANBY are incorrect in their record and are now
corrected by BMC PTF BPM7996 and discussed in their APAR BAM7806.
17. APAR PO56039 for MQSeries V5.2 for OS/390 corrects values in many
SMF 116 fields: WQMAXLAT, WAMINLAT, WQTOTLAT, WQGETMAX, WQGETMIN,
WQPUTMAX, and WQPUTMIN are documented as incorrect.
16. SMF Writer design cannot handle normal bursts of SMF data, for
example when a step with many dynamic allocations ends. These
bursts overrun the SMF buffers, causing loss of SMF data.
A specific, daily example:
ENDEAVOR, an programming development system used by hundreds of
programmers allocates many DDs dynamically for each use. In one
day, from 9am to 3pm, there were 2,548,964 DD segments that were
written in 1,735 type 30 subtype 4 SMF records:
15:06:48.78 First of 1735 Step Termination 30-4 Records
15:06:50.54 Last of 1735 Step Termination 30-4 Records
54 MegaBytes in 1.76 Elapsed Seconds ==> 30 MB/sec
Then 0.70 Elapsed Seconds from end of step SMF to start of job end
15:06:51.24 First Job Record SMFTIME
15:06:52.52 Last of 1735 Job Records SMFTIME
54 MegaBytes in 1.28 Elapsed Seconds ==> 42 MB/sec
Total: 108 MegaBytes in 3.74 Elapsed Seconds ==> 30 MB/sec.
This is far above the sustainable write-to-VSAM-file rate of the
SMF writer with current DASD:
One 3GB volume is filled every 20 minutes ==> 2.5 MB/sec max
And IEE986E messages shows five SMF buffer expansions in 15:06:52.
The SMF Buffer is not designed to absorb 100 MB bursts of data;
not only due to its small maximum size (which itself requires a
PTF to allow the 128MB maximum), but the intrinsic serial design
of the SMF writer causes potential data loss with every buffer
expansion (eight 8MB expansions from 8MB to 128MB), because the
writer cannot accept new records during buffer expansions.
The case of the "multi-DD" step SMF burst occurs randomly, but
more frequent (and often larger) SMF bursts occur at interval
end when there are many CICS regions recording statistics data,
and synchronized interval recording is critical for analysis
across multiple address spaces, increasing the need for redesign.
Loss of SMF data records is both an accounting exposure, and a
security exposure, since any SMF records can be lost during the
buffer expansions to handle these bursts. I believe that the
integrity of the SMF file is a prerequisite for OS/390 being
designated as a B2/C2 DOD Secure operating system.
The SMF Writer clearly must be redesigned to handle these bursts!
Many of the companies in IBM's Gold Consultant program not only
depend on SMF data being written, but recommend z/OS systems
because of the SMF integrity in recording security accesses, and
for performance and capacity measurement, all of which are
threatened by SMF data loss. I would like to propose a telephone
conference call to discuss how members of that group can provide
resources and assistance to IBM to redesign the SMF writer to
handle the current and expected SMF data without data loss.
15. CMF type 70 does not contain the ICF flag, until one or more of
BPM7293, BPM7307, BPM7308, and/or BPM7309 are installed.
Dec 16, 2001.
15. APAR OW52226 for JES3 Only, type 30 fields SMF30PTM and SMF30TPR
(MXG variables TAPNMNTS and TAPSMNTS) do NOT count JES3 dynamically
allocated tape mounts. The APAR is "CLOSED FIN", Fixed-in-Next
(i.e., the next time IBM has to update that code for some other
reason).
14. APAR OW51353 corrects defective VVDS after OW50528. The impact was
defective type 60 SMF records, and IBM still refuses to provide the
VVDS DSECTS, claiming they are "object code only".
13. APAR OW45447 corrects an IEC027I 737-24 ABEND on CONFIG-002 DD; the
IBM error occurs when non-SMS and SMS managed datasets were in the
//CONFIG DD concatenation. Reversing the order so that the non-SMS
data set was first avoided the ABEND, but MXG expects its CONFIGV8
member to be last, because it is the last CONFIG option that is
used by SAS. Previously, CONFIGV6 for V6 had to be last so that it
would override the SAS default for MEMSIZE, but since SAS V8 cannot
have a MEMSIZE parameter in //CONFIG members, the need for MXG's
CONFIGV8 member to be last may not be absolute. I'm checking.
12. APARs OW50579 and OW50837 are marked "VARIOUS RMF PROBLEMS" and
both fix a number of problems with RMF I and RMF III reports and
ABENDs in both RMF Monitor I and Monitor III sessions.
11. SMF 30 records with ALOCTIME earlier than INITTIME (SMF30AST/30SIT)
can occur even though it is physically impossible; APAR OW50134
acknowledges the problem, but that apar is "FIN", Fixed in Next
(i.e., next 18 months in a new release).
10. SMF 42 subtypes 15 thru 19 were not written after APAR OW47863 was
installed; new APAR OW51179 corrects the error.
9. SMF VBS records with LRECL greater than 32760 have been found
beginning with OS/390 R2.7, and IBM has acknowledged their error;
SMF records can have a maximum LRECL of 32760 (max data LENGTH of
32756). APAR OW51139 for SMF Type 8 records, OW51146 for SMF Type
90 Subtype 32 records. These records cannot be read with SAS
V6-V8, so you must use IFASMFDP to copy/delete if they are found.
8. With OS/390 R2.10 and DFSMS/rmm, problems with SAS data sets on
tape required IBM APAR OW49577 for RMM.
7. A user noted that IBM's IXGRPT1 utility report from TYPE88 had the
incorrect values for SMF88ETT and SMF88EO, but MXG was correct.
6. APAR OW49920 reports overlays caused when program objects in PDSEs
are LLA-managed and the object gets staged to VLF.
5. APAR O49773 reports missing RMF Cache data from Shark D/T 2105
after PTF UW77972 is installed.
4. Information APAR II12079 discusses lots of FTP errors and issues,
and reasons why SMF 118 TCP records might not be written for FTP
Client and/or Server: there are two parameters in PROFILE.TCPIP
dataset, SMFCONFIG and SMFPARMS, documented in IP Configuration
Guide, Chapter 3. SMFCONFIG FTPCLIENT is shown as an example to
create the client record.
3. APAR OW50084 corrects non-writing of SMF 79 subtype 1,2, or 5 under
Goal Mode if DOMAIN parameter is specified for ASD/ARD/ASRM, errors
if CPs are varied ON-line and was OFF-line at the end of interval,
and many errors in RMF reports.
2. APAR OW41696 corrects blank SMF30EXN/OMVSEXNP program name field in
USS/OMVS type 30 SMF records.
1. APAR OW38842 corrects Accounting Fields in SMF Type 30 from USS
(unix system services, a/k/a OMVS) tasks. accounting information
was taken from the calling address space instead of the address
space associated with the passed ASSB.
V. IMS Technical Notes.
VI. SAS Technical Notes.
11. MOVE TO 41. 82BA62 Fixed Windows error:
FATAL ERROR: WRCODE=80000805, MODULE='wtdelet': Unexpected return
from vtswtch().
10. A note on case sensitivity of SAS under unix.
Under unix operating systems, a file named x.y is different that a
file named X.y, but when SAS executes under unix, its sensitivity
to case depends on the syntax. This note may not be perfect:
- DDNAMEs and DATASET names are case insensitive in SAS statements
but the created unix file names will be lower cased. SAS will
read x.sas7bdat with SET X; or with set x; in your unix program.
- Stored variable names are in the case of their first instance in
the source program, so if you type the variable name in lower
case it will be stored as a lower case variable name. Thereafter
all references to that variable name are case insensitive
- External files, like MXG Source Directory members, must be lower
cased and end with .sas, so that the %INCLUDE SOURCLIB(MEMBER);
syntax can be transparently used across all platforms.
- MXG intends to always create only upper case names, so as to
avoid the mess created by those unix idiots that thought case
sensitivity was a good idea. Life is complicated enough just to
spell things right; getting the right case is asking too much.
9. An interesting, but probably not generally important note, mostly
about square brackets. The beautiful hex dump created by the LIST
statement (or by input errors like invalid data), interprets some
hex characters differently in the CHAR line of the dump than in the
PUT= of the input of that hex character. The LIST option checks
each character with the C isprint function to test its printability,
but isprint's definition of 'printable' is platform dependent; on
MVS, isprint('[') or (']') is considered unprintable and returns a
zero, so the CHAR line prints a dot when it sees 'AD'x/'BD'x hex.
But when you PUT a character variable, there is no printability test
and so it will be displayed in batch job's logs based on a table in
????????, while under TSO, the character you see in a PUT statement
is controlled by the translate table in the emulator that owns your
screen.
If square brackets do not display on your ISPF terminal, you need to
have your VTAM SysProg set PROGRAM SYMBOLS ON in the VTAM Terminal
Definition for your terminal.
8. MXG Software and SAS Versions that you can use.
a. Independent of which MXG Version you are using:
We STRONGLY RECOMMEND that you ONLY execute MXG Software under V8:
SAS V8.2 TS2M0, plus, for MVS-OS/390-z/OS, with the 82BX03 HOTFIX
Bundle (that includes critical 82BA57 fix).
Original 82BX01, changed to 82BX03 Oct 2002.
but, if you still must stay in the dark ages, there are no reported
MXG errors using SAS V6.09e at TS-475 (with SHARK, check SAS notes).
There were errors in releases of V8 prior to SAS V8.2 that have been
fixed in the V8.2+82BX01 HOTFIX; you'll be wasting your time and my
time trying to use any earlier releases of V8 SAS.
b. MXG will exploit new V8 enhancements when MXG is executed under V8,
but you do not need a new version of MXG when you install a new SAS
release for your daily runs. However, for some new data sources,
execution under V8 is required: Websphere EE SMF type 120 records
contain DBCS Unicode characters that were not supported until V8,
and MXG TYPEWWW WebLog support needs 32000-byte-length-character
variables (which it fills mostly with blanks, which disappear with
compression!).
Note it is only V8-greater-than-200-byte-character-variables that I
use in MXG; there are no long-variable-names, and since I only see
complication and confusion with long-variable-names, I do not now
expect to have MXG variable names longer than eight characters.
However, if alias names for variables ever exist in SAS, I might
want to change my mind. I could use the variable's LABEL as an
alias name, so you could access it either way, and I would have
a third alias with the original IBM field name as variable name!
7. An increase in CPU time between V6 and V8 in a unique MXG job was
found by SAS to be in its keeping track of the count and location of
missing value calculations, and in this special case, the NOSTMTID
SAS option significantly reduced the CPU time. The site had used
IMACEXCL to keep only 7 variables in CICSTRAN, keeping no datetime
variables, which caused latent Y2K protection code to be executed
because those datetime variables were now missing! The added CPU
time was not due to the missing value calculations themself, but
rather the calls to keep count; with OPTIONS NOSTMTID that counting
is bypassed, no counts are kept, and significant CPU reduction in
this unique case. However, tests with NOSTMTID enabled showed a
small 0.1%-0.5% increase in CPU time, so it does not seem worth
enabling by default. And SAS will revise the count-calls in SAS
Version 9, now that they have diagnosed the problem. SN-004513.
10Apr02 update: New UTILEXCL logic to create IMACEXCL completely
eliminate any need for NOSTMTID, as previously expected; but here
are the CPU times of what happened:
V18.18 98.77 Original V609 Run, before problem.
V18.18 229.59 Run with V8.2 that uncovered the problem.
V18.18 105.31 Run with V8.2 with DSOPTIONS=NOSTMTID
V19.19 82.29 Now, using IMACEXCL built by UTILEXCL
6. SAS usage note SN-004513 is an Outstanding Problem that discusses
possible increased CPU time with Version 8 when missing values are
passed to functions; the counting and keeping track of which line
number had missing value calculations, and how many, is apparently
expensive in V8; the note suggests OPTIONS DSOPTIONS=NOSTMTID;
which we are testing. The specific problem was a heavily tailored
CICSTRAN dataset in which IMACEXCL was used to only input 6 fields,
so all of the other variables were missing, which is not normal.
In this case, the option was very significant; the 217 CPU seconds
with out the option dropped to 97 seconds when it was enabled to
create 16 million observations in CICSTRAN.
5. Should you use //SORTWKnn DDs or Dynamic Allocation for SAS Sorts?
MXG has always defaulted to having actual //SORTWK01 -> //SORTWK03
DD statements in its JCL examples. Here's why:
a. There is a limit of 6 Sort Work Areas when they are dynamically
allocated, so if you need more than 6 work units, pre-allocating
//SORTWKnn DDs is the only way to get more space and/or units.
Chuck Hopf uses 64 sort work DDs in his monster ASUMUOW job.
b. If a VIEW or a Sequential-format SAS dataset is involved in the
SORT, SAS can't determine the size to pass to the sort, so use of
pre-allocation can ensure enough work space is allocated.
c. Never use RLSE on //SORTWKnn DD statements, unless only one SORT
is to occur. In BUILDPDB, there are multiple sorts, some large
some small, and some large again, and with RLSE that third sort
can fail with a B37 when you RLSEd the space that somebody else
has got now!
d. The Sort Work DDs must be named //SORTWK01 and be in order, i.e.
SORTWK01, SORTWK02, ... SORTWKnn.
e. The CONTIG parameter is no longer required by sort packages, and
it can delay or even prevent allocation if that large chunk of
work space is not currently available on DASD. Sort benchmarks
have not shown a significant difference with/without CONTIG now.
f. But if you want to know the actual allocation algorithm for SAS
sort work space allocation:
-Host Sort interface is chosen to be used (either by SORTPGM=HOST
or the size of the file to be sorted is greater than the value of
SORTCUTP when SORTPGM=BEST), then
-SAS checks TIOT to see if SORTWK01 is allocated, uses if found.
-Otherwise, check to see if SORTWKDD value (see below) has been
previously allocated, and if the space is sufficient for the
current sort to be to be performed.
-If that space is not sufficient then DEALLOC if present.
-Check if DYNALLOC option is set. If it is, SAS will not do the
allocate but instead lets the sort package dynamically allocate
its sort work area.
-If NODYNALOC is set, then SAS will now allocate the sort work
space:
-SAS DD names are based on the SORTWKDD= option's value, which
defaults to SASS, so they are of the form SASSWK01 to SASSWKnn
where nn is the value of the SORTWKNO= options (and it is hard
limited to six).
-The size of sort work space allocated is is computed as (number
of OBS)*(RECLEN)*(2).
-If the size cannot be determined (VIEW or SEQ Format) then
SORTSIZE= option is used to determine sort work space to
allocate.
-Allocation is based on SORTUNIT (TRKS/CYLS/BLKS), SORTDEV
(3390/3380/SYSDA...). SAS still adds the CONTIG parameter.
-If there is a DAIRFAILure message it is written to the SASLOG
indicating that a dynamic allocation failed allocating that
specific DD, but SAS will continue to attempt the SORT.
SAS Technical Support provided assistance in writing item f.
Dec 12, 2001.
4. Errors when reading a V6.09-built SAS dataset with SAS V8:
LIBRARY SPIN IS NOT IN A VALID FORMAT FOR ACCESS METHOD SASE7.
is probably fixed in Hot Fix 82BA57, below, but is circumvented with
LIBNAME SPIN V609;
that tells SAS V8 to use the V609 engine to read that DDNAME.
This error first occurred after maintenance for IBM SHARK DASD; one
one reporting site was backlevel at SAS V8.0, and they also had an
ERROR FORMAT $MGTRAN NOT FOUND
when V8.0 tried to use a //LIBRARY format library that had been
created by V6.09. Creating a new format library with V8.0 was the
best solution, but a LIBNAME LIBRARY V609; statement circumvented
the error. But, once again, you really need to be at SAS V8.2 with
all Hot Fixes! First: Nov 30, 2001. Last: Jan 10, 2002.
Update July 12, 2002: Either of these error messages:
LIBRARY SPIN IS NOT IN A VALID FORMAT FOR ACCESS METHOD SASE7.
LIBRARY SPIN IS NOT IN A VALID FORMAT FOR ACCESS METHOD SASEB.
are now documented in SAS NOTE SN6447; these messages can occur if
DISP=MOD is coded for an existing SAS Data Library, if you have
installed any of these hot fixes: 81BA50, 82BA57, 82BX01
Their temporary circumvention is to use DISP=NEW, but I will update
this note after I discuss further - SAS needs to support DISP=MOD.
3. OS/390 Hot FIX 82BA57 for V8.2 fixes many I/O & multi-volume errors
(and Hot FIX 81BA50 is for V8.1). SAS note SN-05642 discusses and it
lists these SAS notes that are associated with this defect:
895 2674 2881 4229 4916 5004.
Problems included non-read VBS records after error, multiple mounts,
BY variable position errors, S0C1 in SORT step, B37-04 on SYSOUT DD
with a PROC PRINTTO, U315/U317/S0C4, etc.
Nov 20, 2001.
2. A minor difference in values between ASCII and EBCDIC SAS numeric
variables that are stored in less than eight bytes is documented.
There is no error, just differences in the way numeric values are
stored on different hardware platforms and/or operating systems!
MXG 19.11 now uses LENGTH DEFAULT=&MXGLEN and &MXGBYLN (See Change
19.272) to set the stored length of most numeric variables, where
MXGLEN is 5 on EBCDIC and 6 on ASCII, which both saves disk space,
and keeps full resolution of fields input from 4 bytes. Only if the
exact value of a large-value variable is needed is LENGTH 8 used;
datetimestamp and byte variables are length 8.
Because SAS uses floating point internal storage, ASCII systems
require a minimum of 3 bytes to store a one-byte numeric, while
EBCDIC systems only require 2 bytes.
Previous EBCDIC truncation measurements when 'FFFFFFFF'x was INPUT
as PIB4 and stored in 4 bytes had shown a maximum loss of 255, but
these new ASCII measurments under WINDOWS show a maximum loss due to
truncation of 247.
Hex Value INPUT PIB4. ASCII INPUT PIB4. EBCDIC LENGTH
FFFFFFFFx 4294967295 4294967295 8
FFFFFFFFx 4294967295 4294967295 7
FFFFFFFFx 4294967295 4294967295 6
FFFFFFFFx 4294967288 4294967295 5
Truncate loss: 7 0 5
FFFFFFFFx 4294965248 4294967040 4
Truncate loss: 2047 255 4
This shows that LENGTH=5 for MVS and LENGTH=6 for ASCII platforms
will store all fields input as PIB4 without any truncation.
Originally, Nov 14, 2001, I wrote:
There's really nothing to fix here, but if you ever need the
change the length of an MXG numeric variable, you can override
the MXG code by adding a statement: LENGTH variable 8; in your
EXdddddd "exit member" in your tailoring library. SAS uses the
last instance of a LENGTH statement, and the Dataset Exit
member's code is always seen after the MXG LENGTH statements!
Nov 29, I realized I can easily externalize the MXG default value
using DEFAULT=&MXGLEN in place of DEFAULT=&MXGLEN in all MXG source
and preserve the existing value by using %LET MXGLEN=4; in the MXG
initialization, and then, if you need full resolution, you can have
it by using %LET MXGLEN=n; as your first //SYSIN statement in the
MXG jobs that create the datasets. Implemented in Change 19.272.
If you use a PUT variable= ; statement in a SAS program that is
creating and then storing that variable in 4-bytes, the output of
the PUT will be the full 8-byte value, but a PROC PRINT of the
dataset will have the (truncated) 4-byte value; SAS uses 8-byte
virtual storage internally for all numerics; it is only when the
variable's value is OUTPUT to a dataset that its stored length is
set by the LENGTH= statement.
1. SN-002861 is a SAS V8.0 and V8.1 only correction; the error is fixed
in SAS V8.2. Tape Format libraries on disk which encountered B37
out of space conditions require zap z8002861 or Z8012861, and these
zaps also correct errors in SN-004787 with S837-08 and multiple tape
volume datasets. 0C4's may also occur, but again, only if you are
still running 8.0 or 8.1.
VII. CICS Technical Notes.
1. APAR AQ61844, about two years old, for CICS 4.1, reduced CPU time
in LE and COBOL environment's initialization and termination with
modules ceevgtsi, igzeini, and igezetrm all active.
VIII. Windows NT Technical Notes.
IX. Incompatibilities and Installation of MXG 19.19.
1. Incompatibilities introduced in MXG 19.19 (since MXG 18.18):
a- Landmark data for DB2, IMS, and MVS now have datetime variables
converted to local time zone by MXG; previously the datetimes
were incorrectly left in GMT. This could affect reports if you
use TYPETMDB, TYPETIMS, or TYPETMV2. The Landmark CICS data was
not corrected. See documentation in Change 19.288.
2. Installation and re-installation procedures are described in detail
in member INSTALL (which also lists common Error/Warning messages a
new user might encounter), and sample JCL is in member JCLINSTL.
X. Online Documentation of MXG Software.
MXG Documentation is now described in member DOCUMENT.
XI. Changes Log
--------------------------Changes Log---------------------------------
You MUST read each Change description to determine if a Change will
impact your site. All changes have been made in this MXG Library.
Member CHANGES always identifies the actual version and release of
MXG Software that is contained in that library.
The CHANGES selection on our homepage at http://www.MXG.com
is always the most current information on MXG Software status,
and is frequently updated.
Important changes are also posted to the MXG-L ListServer, which is
also described by a selection on the homepage. Please subscribe.
The actual code implementation of some changes in MXG SOURCLIB may be
different than described in the change text (which might have printed
only the critical part of the correction that need be made by users).
Scan each source member named in any impacting change for any comments
at the beginning of the member for additional documentation, since the
documentation of new datasets, variables, validation status, and notes,
are often found in comments in the source members.
Alphabetical list of important changes after MXG 18.18 now in MXG 19.07:
Dataset/
Member Change Description
See Member CHANGES or CHANGESS in your MXG Source Library, or
on the homepage www.mxg.com.
Inverse chronological list of all Changes:
Changes 19.288 thru 19.001 are contained in member CHANGES.