290 likes | 484 Views
CMG 2004. MXG Update 1. IFAs a/k/a zAAPs – what they are, where’s CPU time captured? 2.MXG 22.08 required for SAS Version 9.1 – other SAS issues. 3. ASMTAPES/MXGTMNT – Allocation Recovery Monitor 4. BUILDPDB MVS runtime elongation if ANY output to tape 5. MXG 22.04 required for CICS/TS 2.3
E N D
CMG 2004 MXG Update 1. IFAs a/k/a zAAPs – what they are, where’s CPU time captured? 2.MXG 22.08 required for SAS Version 9.1 – other SAS issues. 3. ASMTAPES/MXGTMNT – Allocation Recovery Monitor 4. BUILDPDB MVS runtime elongation if ANY output to tape 5. MXG 22.04 required for CICS/TS 2.3 6. MXG 22.06 required for DB2 Parallel 7. If you are using IRD, you must install MXG 22.12 or later: 8. Still OS/390? MXG 22.07 required for z890, 21.04 for z990. 9. MXG/SAS V9.1 benchmarks Windows XP, Linux RH8, z/OS 1.4.1 10. Running out of memory “Below the Bar” H. W. "Barry" Merrill, PhD Merrill Consultants Dallas, TEXAS www.mxg.com Wednesday, December 8, 2004
1. What are IFA's, a/k/a zAAPs? • For z/os on z890s and z990s, IFA was the internal name of zAAPs: • zAAPs -zOS Attached Assist Processors • all of the RMF/SMF field names use IFA • zAAP is the marketing name for these engines. • IFAs are engines that execute only Java code • are not included in MSU capacity • can add new Java applications, or offload current Java to IFAs • can increase hardware capacity without increase software costs. • can force all Java workload to execute only on IFA engines • maximizing the offload of work from your CP engines, or • can let Java work execute on CPs when they are not busy: • can let it run at the priority of the Service Class, or • can be run even lower than discretionary. • keywords of IFACROSSOVER and IFAHONORPRIORITY let you choose
Where does IFA CPU time show up: • What's really important about these details on CPU measurement? • That it exists at all: IBM has been extremely pro-active, at • NYC SHARE meeting in NYC in August, provided much of these data • well before z/OS 1.6 delivery - prereq to use the IFA engines. • Below info taken from those SHARE presentations plus follow up • from IBMers, with additional valuable input from Don Deese. • Service Units: • In all records that contain Service Units, the TCB Service Units • that formerly had only the service units due to TCB time on CP • engines now includes SUs from both IFAs and on CPs. If you use • Service Units for billing, your billing units may be increased • when Java work runs on IFAs. • Why? Because WLM manages based on service units; an all-Java work • could have zero TCB service units, could be using 100% of IFAs, • and WLM wouldn't know, if IFA service wasn't included in TCB.
SMF 30. • Existing CPUTCBTM TCB CPU time does NOT include IFA CPU time. • existing billing based on TCB CPU time will not change • (but CPU time on IFAs will not be recovered until you change • your billing code). • New CPUIFATM is the CPU time on IFAs • you can charge IFA CPU time at different rate than CP CPU time • or could add the CPUIFATM to the total CP CPU time in CPUTM • and charge both CPU times at the same rate. • On z990s the CPs and IFAs run at the same speed • On z890s the IFAs run at full speed (365MIPS) while the CPs • run at 28 different "speeds", as low as 27MIPS: • IFA CPU time are normalized back to CP speed of z890 model. • MXG will NOT add the CPUIFATM into the existing CPUTM variable.
New variable CPUIFETM (‘eligible') contains the CP CPU time • that was executed on a CP, but that was elibible to run on IFA. • (CPUIFETM is included in CPUTCBTM). • With z/OS 1.6 and Java SDK V1.4 CPUIFETM measures exactly the • Java CP CPU time that could be offloaded to an IFA. • The zAAP Projection Tool (WSC White Paper WP100431) can be used • now on z/OS 1.4+ to estimate current CP CPU time for Java apps. • Bad news: While IFA CPU time is available in the SMF 30 record, • CPUCOEFF, SU_SEC, and R723NFFI (normalization) factors were NOT, • so without a table lookup, you can't back out IFA Service Units • from the total IFA+CP Service Units in CPUUNITS in type 30s, • new APAR OA09118 adds those and MXG supports. • Maybe bad news: There are concerns as to the repeatability of • CPU metrics, especially since Java work can run on either IFAs • or on CPs when IFACROSSOVER=YES is specified. • Additional concerns for the repeatability of the normalization • factors on z890s. But data from real sites running real Java • work will answer the magnitude of these concerns. • The good news: since so few sites are actually running any Java • on their current z/OS systems, future concerns won't impact the • current chargeback, and you can benchmark and measure your work • repeatability before IFAs are placed in production!
SMF 70. • TYPE70 MVS System data flags each LCPU, new IFATYPnn variable • indicates whether an engine is a CP or an IFA. • PCTCPUBY calculation and related capacity metrics still based • ONLY on the CP engines, as has always been the case. • PCTIFABY calculation and new IFA capacity metrics are added • to TYPE70 dataset so IFA capacity can be separately tracked. • IBM RMF CPU Activity report has separate lines for CPs and IFAs • TYPE70PR PR/SM LPAR data still has only 'ICF' in SMF70CIN for • ICFs, IFLs (Linux) or for IFA engines. • But RMF development has acknowledged that true identification • of each engine type is an urgent requirement. • But for now, can only count how many CPs and how many "others" • engines are available for each LPAR in the PR/SM dataset.
SMF 72. • R723CCPU/CPUUNITS will contain sum of IFA and CP Service Units! • CPUTCBTM is calculated from R723CCPU and SU_SEC and CPUCOEFF. • BUT: MXG variable CPUTCBTM will still contain ONLY the CP CPU • so existing CP capacity measurements, capture ratio unchanged • by subtracting new R73IFAT/CPUIFATM variable (IFA CPU time) • from R723CCPU/CPUUNITS when creating MXG's CPUTCBTM. • R723IFCT/CPUIFETM contains CP CPU time that was IFA-eligible. • The bad news: The RMF Workload Report shows sum of CP+IFA time • in "TCB" line, which will no longer match the CPUTCBTM in MXG. • The RMF report does have a new line with "IFA" CPU time with • CPUIFATM, and new APPL% CP, APPL% IFACP, and APPL% IFA values. • The R723IFAT and R723IFCT times are actually converted from IFA • service units using the new R723NFFI normalization factor.
SDSF display. • JOB-level data for CPU% includes both CP and IFA CPU percent • but a new IFA CPU column will eventually be added. • The Bad News: The CPU percent busy field in the top line of the • display DOES (incorrectly?) include IFAs in the denominator: • a two-CP one-IFA system with 100% in both CPs and 0% in IFA • displays CPU Busy of 66% on SDSF, because the capacity was • based on three total engines, not the two CP engines. • Unrelated: it was just observed that the JOB %CPU in SDFS • includes "Initiator" CPU time (CPITCBTM/CPISRBTH see ADOC30), • CPU time before your program actually starts executing, so • can see CPU time while watching a job that ultimately never • records any CPU time in the IEF374I messages!
IEF374I step end and IEF376I job end joblog messages contain • sum of CP and IFA CPU time, but IBM intends to add a IFA value, • or more likely create a new IEFnnnI message with IFA CPU time. • CICS TASCPUTM already includes J9 Java TCB's CPU time, • available separately in the J9CPUTTM variable, • but that is the total Java CPU time for the transaction. • Won't tell how much was CP versus how much was IFA time. • You will have the SMF 30s for each CICS region, and interval • records, to measure how much time is CP, IFA, and IFA eligible • at least at the region level. • May have to use those percentages in your chargeback code. • IMS Message CPU time - unconfirmed, but it is expected that the • IMS 07 log record contains both TCB and IFA CPU time.
2. Support for SAS Version 9.1 • MXG has been tested under SAS Version 9.1 TSM0 on z/OS 1.4, • Windows 2000, Windows XP, and Red Hat Linux, with no • significant errors nor any real problems. • Most importantly, there are no data incompatibilities • between V8 and V9. • Data libraries and catalogs built with V8 can be read and written • with SAS V9, and libraries and catalogs built with V9 can be read • and written with SAS V8. • Remove MEMSIZE=64M from your CONFIG member • (use MXG CONFIGV8 or CONFIGV9 member); • REGION=0M or 80M controls SAS virtual storage in V8+. • V8+: Use SAS member (BATCH) instead of (BATCHXA) • in your //CONFIG concatenation in your JCL, • V9+ Use ENTRY=SAS instead of ENTRY=SASHOST • (or use MXGSASV8 or MXGSASV9 JCL procedure example).
2. Support for SAS Version 8.2, 9.1, 9.1.2, and 9.1.3 • The good news: • Most importantly, there are no data incompatibilities • between V8 and V9. • Data libraries and catalogs built with V8 can be read and written • with SAS V9, and libraries and catalogs built with V9 can be read • and written with SAS V8. • Remove MEMSIZE=64M from your CONFIG member • (use MXG CONFIGV9 member); • REGION=0M or 80M controls SAS virtual storage in V8+. • V8+: Use SAS member (BATCH) instead of (BATCHXA) • in your //CONFIG concatenation in your JCL, • V9+ Use ENTRY=SAS instead of ENTRY=SASHOST • (or use MXGSASV8 or MXGSASV9 JCL procedure example).
The bad news: • MXG 22.08 required for safe execution with SAS V9.1.2 or V9.1.3. • While MXG 22.07 had critical revisions for SAS 9.1.2, more design • changes were discovered in V9.1.2 that required more MXG changes. • Plus, errors in SAS ToolKit used to update Syncsort's add-on product PROC SYNCSORT for V9, corrupts INFORMATs, caused fatal errors in BUILDPDB: • I removed all MXG INFORMAT statements faster than they could • examine their error, so you can use PROC SYNCSORT with MXG • even though they have not fixed their error. • Errors are in PROC SYNCSORT add-on (prints "PROC SYNCSORT" on • SAS log - no errors with the Syncsort SORT product itself. • Critical actions for you to run MXG with SAS V9.1+: • Install MXG 22.08, use MXGSASV9 and CONFIGV9 from 22.08, and • Run UTILS2ER utility against all of your SAS programs to see • if any lines conflict with S2=72 option that replaced S2=0 • option set by MXG previously.
Specific Changes related to SAS V9.1.2 and MXG execution: • CONFIGV9: • NOTHREADS required for SAS V9.1.2, fixed in 9.1.3. Change 22.207. • SAS Note SN-12943 reports incorrect results, no error message: • PROC MEANS, SUMMARY, REPORT, TABULATE • Only on "MVS", only if threading is used. V9 default is THREADS • While fixed in 9.1.3, I chose to force NOTHREADS in CONFIGV9. • Can use OPTIONS=THREADS with 9.1.3 on // EXEC to change. • CONFIGV9: • NLSCOMPATMODE required: SAS V9 changed default to NONLSCOMPATMODE • Only on "MVS" (thus far!), doesn't fail if LOCALE is ENGLISH/blank • But with LOCALE=GERMAN_GERMANY or other non-blank, or non-ENGLISH • every @ symbol causes an error at compile time. • Extensive discussion in text of Change 22.129 for NLS and LOCALE.
CONFIGV9: • S2=0 option now required; MXG previously used S2=72 in CONFIGxx. • Only on "MVS". Extensive discussion in Change 22.123. • S2 sets line size of source read by %INCLUDE or AUTOCALL. • V9 MVS SASMACRO library was changed from RECFM=FB to RECFM=VB • -no standard for line size of SAS-provided %macro text • -new macros were written by ASCII folks, line length 255 • -Rather than make the authors correct, RECFM changed to VB. • BUT: RECFM VB has entirely different meaning for S2 than FB. • S2=72 FB ==> read only first 72. VB: ==> START IN 72!!!! • MXG had always specified S2=72 to protect you from line numbers • S2=0 ==> look at last 8 columns to see if line numbers exist • All MXG code is only 72 positions, so S2=0 is no-risk to MXG. • BUT: If you have SAS code with mixed blanks and numbers • S2=0 will cause your code to syntax error. • So: New UTILS2ER utility will read all of your source libraries • and identify any exposures in your SAS programs.
CONFIGV9: • V6SEQ may still be required with SAS V9.1, V9.1.2 • Only on "MVS". SN-012437 and Change 22.108 discuss. • SAS V9.1 and V9.1.2 create corrupted and unreadable datasets • with no error at create time, and data is unrecoverable, • if V7SEQ, V8SEQ, or V9SEQ are used. • SAS Hot Fix in SN-012437 does correct the error for V9.1/9.1.2 • BUT: I can't guarantee you have that hot-fix installed, so • MXG SEQENGINE default was again set back to V6SEQ in 22.05. • But: V6SEQ failed with long-length variables, so • Change 22.108 shortened all from-MVS variables. • MXG has had numerous iterations on SEQENGINE. • Mostly because unnecessary compress was done.
MXGSASV9: • MVS JCL Example has new symbolics for NLS/LOCALE options. • XX='EN' - Default Language Value (ENGLISH) • YY='W0' - Default Encode Value (USA) • 'DEW3' is for most GERMAN, but 'DEWB' is for SWIZTERLAND. • You must look at the SAS JCL proc built by your SAS installer • to find the correct XX and YY values, and then set them as • your MXGSASV9 JCL Procedure defaults. • //CONFIG DD DSN= ... CNTL(BAT&YY.) • //SASAUTOS DD DSN= ...&YY..AUTOLIB • //SASHELP DD DSN= ...&XX.&YY..SASHELP • // DD DSN= ...EN&YY..SASHELP • //SASMSG DD DSN= ...&XX.&YY..SASMSG • // DD DSN= ...EN&YY..SASMSG • New DD statements for TRNSHLP, ENCDHLP and TMKVSENV were added.
ASCII-execution code change: • EBCDIC character variables INPUT with $VARYING had hex zeros • where they should have had blanks • because of a SAS V9 Design Change in $VARYING informat. • $VARYING always has returned a "raw" $CHAR string that must • be converted if the string is EBCDIC text, using: • INPUT VARIABLE $VARYINGnn. LENTEXT @; • VARIABLE=INPUT(VARIABLE,$EBCDICnn.); • but when LENTEXT was less than nn, the "pad" of '80'x was • found on SAS ASCII platforms, so the statement • VARIABLE=TRANSLATE(VARIABLE,' ','80'x); • was added to translate the unexpected/undocumented '80'x. • Now, also undocumented, in V9, the "pad" of '00'x is returned! • So an additional • VARIABLE=TRANSLATE(VARIABLE,' ','00'x); • had to be added 511 times in 55 members.
3. ASMTAPES/MXGTMNT ML-29 - ALLOCATION RECOVERY MONITOR • Brand new in MXG 21.04 dated Aug 25, 2003, enhanced • ASMTAPES at ML-29: • new SMF subtype created by MXGTMNT • new MXG dataset automatically created PDB.TYPETARC • each Allocation Recovery: Job, How Long Delayed, etc. • a job must wait because there is no tape drive • applies to real and virtual tapes
4. BUILDPDB MVS runtime elongation if ANY output to tape • - like CICSTRAN or DB2ACCT • - or even if output is sequential format on DASD. • - introduced in MXG 19.19, %VGETENG added for RMFINTRV • to test if a //SPIN DD existed. • - no elongation if no tape or sequential format output • - PROC SQL with FROM DICTIONARY.MEMBERS in VGETENG to get • ENGINE of //SPIN, but no WHERE LIBNAME=SPIN clause used; • it read all LIBNAMEs to populate DICTIONARY.MEMBERS • - If CICSTRAN and DB2ACCT both multi-vol on tape, log has: • hh:mm-hh:mm • SMF Opened, read started 14:25 • CICSTRAN Mount-Dismount 5 vols 14:24-16:06 • DB2ACCT Mount-Dismount 2 vols 14:24-15:25 • SMF Closed, read completed 16:12 • VGETENG-remount/read 2 DB2ACCT vols 16:17-16:30 • VGETENG-remount/read 5 CICSTRAN vols 16:40-17:09 • Total Elapsed time: 164 minutes with re-read. • VGETENG: wasted 52 minutes mounting and rereading • - And PROC SQL prints NO messages about CICSTRAN/DB2ACCT • - Only clue to elongation are those extra tape mounts. • - Fixed in MXG 22.01+, can circumvent with FREE=CLOSE
5. MXG 22.04 is required to support CICS/TS 2.3 SMF records • MXG 21.04 supported UTILEXCL to read CICS/TS 2.3 SMF 110s. • If you used UTILEXCL to read your CICS/TS 2.3 Dictionary to • create IMACEXCL, the IMACEXCL correctly read SMF 110 data. • You MUST use UTILEXCL if there were EXCLUDEd fields. • You SHOULD use UTILEXCL always as it only outputs the variables • that exist in your current release(s) of CICS. • But if you read SMF 110 records with MXG 21.04 thru 22.03, and • all fields were present, the TASCPUTM and many other variables • were completely wrong, and there were no error messages. • All my CICS/TS 2.3 test data had EXCLUDEd fields!
6. MXG 22.06 is required for DB2 Parallel • CPU time for DB2 Parallel Trans was not output • (i.e., lost, could be very large) in DB2ACCT. • Code in MXG Exit Members EXDB2ACC/EXDB2ACP/EXDB2ACB • deleted all obs with DB2PARTY=‘P’, which was wrong • because those obs contain the DB2TCBTM for parallel events. • New DB2PARTY=‘R’ for the Roll-Up observations also added. • Extensive DB2 Technical Note in Newsletter FORTY-FIVE and • additional documentation in Change 22.121 text.
7. If you are using IRD, you must install MXG 22.12 or later: • Full Support for IRD (Intelligent Resource Director) in all • CPU-related datasets. IRD support was incremental in MXG: • Datasets When MXG Version Change • ASUM70PR/ASUMCEC Sep 22, 2003 21.05 21.170 • TYPE70PR Mar 11, 2004 22.01 22.011 • TYPE70,RMFINTRV Dec 2, 2002 22.12 22.305 • PCTCPUBY in TYPE70 and RMFINTRV were wrong in any interval • when IRD varied CPUs offline. • I'm embarrassed, since PCTCPUBY is the second most important variable in all of MXG • (CPUTM for billing is the most important); • This is the first PCTCPUBY error in MXG's TWENTY-YEAR history! • When all engines remained online, however, there was no error.
8. Still OS/390? • MXG 22.07 is required for z890 and 21.04 for z990s. • IBM changed CPUTYPE value • z990 – 2084x • z890 – 2086x • Only impacts MSU variables that MXG had to set via • a table lookup based on CPU TYPE for OS/390. • With z/OS, MSU fields are in the SMF records so there • is no table lookup required.
9. MXG/SAS V9.1 benchmarks Windows XP, Linux RH8, z/OS 1.4: • -Linux and Windows runs on same AMD 1400 1.5GHz, 500MB ram • -z/OS runs on IBM 2064-210. An 842 MB SMF file was used: • -TYPE30 DATA step cost, and PROC SORT TYPE30D (3.4 GB): • Data Step LINUX WINDOWS z/OS • Elapsed Time 4:27.70 7:40.02 11:03.36 • CPU time 3:59.46 3:57.64 5:56.70 • PROC SORT • SORTSIZE DEFAULT 48MB 64MB MAX • Elapsed 15:39.82 28:19.89 12:28.98 • User CPU 5:01.43 4:02.93 5:23.16 • SORTSIZE 200MB 200MB • Elapsed 15:26.10 28:19.89 • User CPU 5:01.02 4:02.93 • SORTSIZE 400MB 400MB • Elapsed 19:12.40 35:05.38 • User CPU 5:02.79 4:15.81
Benchmark observations: • - SAS V9.1 under LINUX significantly outperforms Windows XP • in both the DATA step and in the SORTS. • - SAS V9.1 under z/OS is significantly slower than either • LINUX or Windows, for the DATA step, but PROC SORT on z/OS, • which used SYNCSORT, is better than Windows or Linux. • - but: SYNCSORT on Windows product tests with SAS V9.0 • was significantly better than the V9.0 SAS Internal SORT. • (See Newsletter FORTY benchmarks). • - SYNCSORT on Windows was not available for these runs. • - SORTSIZE on ASCII does impact elapsed time • - elongation occurs if SORTSIZE is too large or too small • - SAS V9.1 defaults were increased from old 2MB default • and seem fine for this 3.5 GB sort.
Benchmark conclusions: • - Real message: the repeatability and reliability of SAS: • - 5-10 minutes to read 1 GB SMF file to create a PDB lib • - 15 minutes to sort a 4 GB dataset • NO MATTER WHAT PLATFORM YOU CHOOSE FOR SAS. • BUT: this is NOT a capacity comparison of these platforms • - running MXG on ASCII requires dedicated hardware • at least during the creation of PDB libraries • neither Windows nor Linux/unix have Workload Manager • BUILDPDB blocks out other users on shared unix platform • - Gives you confidence that MXG and SAS will continue to • measure computer systems, no matter where SAS runs, • - Your MXG and SAS skills are transferable across platforms • - You can keep your job!
10. Running out of storage "Below the bar" is catastrophic: • - Failing system had to be removed from the SYSPLEX • - SYSPLEX recovery took 3 minutes, halting all systems • - Caused by twenty parallel sorts with SYNCSORT • HIPER APAR OA03577 from IBM for RSM/SRM • plus fix from SYNCSORT for z/OS 1.1 release • if you have lots of sorts with DSM enabled. • - Sort jobs fixed 99% of page frames 16MB ==> 2GB "Bar" • RSM failed to detect the page shortage "below the bar", • so SRM did not take any action. • APAR addresses RSM problem so that SRM takes action • SYNCSORT fix limits storage that it fixes. • Only occur with SYNCSORT's global DSM option enabled • - Turning off DSM limits job to the VSCORET parm • - Turning off DSM helps overall system, but • can elongate run times for large sorts.
KELLER WILLIAMS TO BE FEATURED IN THE DECEMBER 8th EPISODE • OF VH1’S NEW WEEKLY SERIES “MY COOLEST YEARS.” • It’s doubtful that high school was the best time of your • life. But there are certainly plenty of entertaining stories • to tell. Don’t miss Keller Williams’ hilarious recollections • of his high school years, and his memories of those “social • survival strategies” that got him through it all. • VH1'S NEW WEEKLY SERIES "MY COOLEST YEARS” RELIVES THE • CHOICES WE MADE IN ORDER TO FIT IN WITH THE JOCKS, THE • CHEERLEADERS, THE GEEKS, THE B-BOYS OR THE METALHEADS -- ONE • CLIQUE AT A TIME. EACH EPISODE FEATURES A MELTING POT OF • ROCKERS, RAPPERS, ACTORS AND GADFLIES PROUDLY SHARING • MEMORIES OF THE FRIENDS THEY PICKED AND THE STUNTS THEY • PULLED IN ORDER TO SURVIVE THE “COOLEST” YEARS OF THEIR • LIVESB • One December 8th, “My Coolest Years” examines The Dirty • Hippies – and includes stories from Keller Williams, Steve-O • of Jackass, Dave Navarro, and many others. The episode airs • at 10/9c PM For more information visit • http://www.vh1.com/shows/dyn/my_coolest_years/85910/episode_about.jhtml.