-
IBM pSeries - page 1
IBM ~ pSeri es Hig h Perf o rm ance Sw it ch Tuni ng an d Debug Gui de Versi on 1.0 April 2005 IB M Systems and Technology Group Clus ter Perf ormance Department Poughk eepsi e, NY ...
-
IBM pSeries - page 2
pshps t unin ggui d ewp0 401 05. doc P a ge 2 Content s 1.0 Introduction..................................................................................................... 4 2.0 Tunables a nd settin gs for switch softwa re ...................................................... 5 2.1 MP I tunab les for Pa ra llel Env ironm ent .................... ...
-
IBM pSeries - page 3
pshps t unin ggui d ewp0 401 05. doc P a ge 3 5.10 MP_PRINTEN V ...................................................................................... 22 5.11 MP_STATIS TICS .................................................................................... 23 5.12 Dropped switch packets ............................................................ ...
-
IBM pSeries - page 4
pshps t unin ggui d ewp0 401 05. doc P a ge 4 1.0 I ntroduc tion This pape r is in tended to hel p you tune and de bu g the p e rf or man ce of the I BM ® pSe ries® High P erform a nce Switch (HPS) o n IBM Cluste r 1600 system s. It is n o t i ntended t o be a com pr ehen s ive gu ide , but rathe r to hel p in initi al tuning a nd debugging of pe ...
-
IBM pSeries - page 5
pshps t unin ggui d ewp0 401 05. doc P a ge 5 2.0 Tunables and settings f or switc h software T o opt i miz e t h e HP S , you c a n s et sh el l va r i a b l es f or P ar a llel E nv ir o n men t M P I - b a s ed w or k l oa ds an d for I P-bas e d w ork lo ads . Th is se cti on rev ie w s th e s hel l v ariab le s th at are m o st o f ten used fo ...
-
IBM pSeries - page 6
pshps t unin ggui d ewp0 401 05. doc P a ge 6 th read , and from wi t hin the MP I/LAPI pol ling code th a t is invoke d when the appl ic ation m akes blo cking MPI call s. MP_POLLING_IN T ERVA L sp e cifie s the num ber o f micro seconds an MPI/LA PI se r vice t hre ad sho uld wait ( slee p ) be fore i t check s whe ther any data prev iously sen t ...
-
IBM pSeries - page 7
pshps t unin ggui d ewp0 401 05. doc P a ge 7 2.1. 5 MP_TASK_AFFINI TY Se tting MP_TA SK_A FFINI TY to SNI te lls parallel ope r ating envi r onmen t (POE) to bind each task to the MCM con taining the HPS adap te r it w ill use, so th at th e ad apte r , CPU, an d mem or y used by any task a re all lo cal to t he sam e MCM. To pre ven t m ulti ple ...
-
IBM pSeries - page 8
pshps t unin ggui d ewp0 401 05. doc P a ge 8 So meti mes M P I-IO is u s ed in an a p plicat ion a s if it we r e ba s ic P OS IX read/writ e, eit h er becau s e the r e i s no nee d for mo re complex re a d/ write patte rns or be cause the a pplication w as pr evi ously h and-o ptimize d t o use POSIX re ad /w rite . In such cas es, i t is o ften ...
-
IBM pSeries - page 9
pshps t unin ggui d ewp0 401 05. doc P a ge 9 rfifosize 0x1000000 receive fifo size False rpoolsize 0x02000000 IP receive pool size True spoolsize 0x02000000 IP send pool size True 3.0 Tunables and setti ngs for AIX 5L Seve r a l se ttings in AI X 5L im pact the pe rfo r m a nc e o f t he H PS. T h e se in clude the IP a nd mem o r y subsy stems. T ...
-
IBM pSeries - page 10
pshps t unin ggui d ewp0 401 05. doc P a ge 10 The ove rhe a d in m aintain ing the file cac he can im pact t he p e rfo r mance of large paralle l app l i c atio n s. M uch o f t he ove r he ad i s ass o ciate d w i th the sync() syst em cal l (b y defau lt, run ev er y mi nu t e fr o m t h e sy n cd daemon). The sy n c() sy st em call sc ans al l ...
-
IBM pSeries - page 11
pshps t unin ggui d ewp0 401 05. doc P a ge 11 3.3. 1 svmon Th e svm on com m a nd provide s in for m ation about the virtual me mo r y usage by the kernel and u s er p r oc es s es i n t h e s ys t e m a t a ny g i v e n t i m e. F or ex a mp l e, t o s e e s y s t e m - w i d e i n f or ma t i o n about th e segme n ts (256MB chunk of virtual m e ...
-
IBM pSeries - page 12
pshps t unin ggui d ewp0 401 05. doc P a ge 12 Pa geSiz e Inu se Pin P gsp Vi rtu al 4KB 448221 3687 2675 449797 16 MB 0 0 0 0 Vs i d E sid T yp e De sc r ip ti on LP a ge Inu se P in P gsp Vi rtu a l 1f187f 11 w ork te xt data BSS he a p - 56789 0 0 56 789 218 a 2 70000000 w ork defaul t shm a t/ mm a p - 33680 0 0 33 680 131893 17 w o rk te xt d ...
-
IBM pSeries - page 13
pshps t unin ggui d ewp0 401 05. doc P a ge 13 statisti cs i n 5-se cond in terv als, wi th t he f irst se t of statisti cs being the statisti cs si nc e the node o r LP AR wa s la s t b oot e d . vm sta t 5 The pi and po of the page group i s the number of 4KB pages re ad f rom and w r itten to the paging devic e b et ween co ns ecuti ve sa m pli ...
-
IBM pSeries - page 14
pshps t unin ggui d ewp0 401 05. doc P a ge 14 adapte r is c onf ig ur ed . Th e vol ume of re serv ation i s proportion a l t o the number o f user w indow s con figured on the HPS a dapte r. A priv ate window i s required f o r each MP I task. He r e i s a f o rm ula t o c al c ul at e t he num be r o f TLPs nee de d by th e HPS ad apte r. I n th ...
-
IBM pSeries - page 15
pshps t unin ggui d ewp0 401 05. doc P a ge 15 3.5 Large pages and IP support One o f the mo st important w ay s to im pr ove I P pe rfor mance on the HPS is to en sur e th a t large pages a re en abled . Lar ge pa ges are re quired to a llocate a n u m b e r of l arge page s which will used b y t h e H P S I P d r iv er a t b oot t i m e. Each sn ...
-
IBM pSeries - page 16
pshps t unin ggui d ewp0 401 05. doc P a ge 16 If yo u have eigh t cards fo r p690 (o r four ca rds fo r p655 ), thi s com ma nd al so indic ates whe ther yo u have fu ll mem o r y bandwid t h. 3.8 Debug set tings i n the AIX 5L kernel The AI X 5L ke rnel has seve ra l de bug se tting s th at a ffect the performan ce of an appl ic ation. To m a ke ...
-
IBM pSeries - page 17
pshps t unin ggui d ewp0 401 05. doc P a ge 17 4.2 LoadLev el er dae mons The LoadLevele r ® da emon s ar e ne eded f or MPI application s using HP S . Ho weve r , you can lowe r the im pact on a parallel a pplic ation by ch anging the de fault se tting s fo r the se d aemon s. You c an lowe r the im pact o f the L oad Levele r daemon s by: • Re ...
-
IBM pSeries - page 18
pshps t unin ggui d ewp0 401 05. doc P a ge 18 SC HEDD_DEBUG = -D_A LWAYS 4.3 S ett i ngs for AIX 5L threads Seve ral v ar iable s hel p you use AIX 5L th r e a ds to tune pe r f orm ance . The se are the recomm ende d in itial se tting s fo r A IX 5L th re a ds wh en using HPS. Set th e m in the /et c/en vironm ent file . A I XTHREAD_SCOP E=S A IX ...
-
IBM pSeries - page 19
pshps t unin ggui d ewp0 401 05. doc P a ge 19 5.0 Debug set tings and dat a collect ion tools S e v er a l d ebu g set t i n gs a nd da t a c ol l e c t i o n t o o ls c a n h el p you de b u g a p er f or ma nc e p r ob l e m o n sy st em s using HPS. Th i s se ction c on tains a sub se t of the mo st c omm on setting c hange s and t o ols . I f ...
-
IBM pSeries - page 20
pshps t unin ggui d ewp0 401 05. doc P a ge 20 5.3 Affini t y L PARs On p690 sy stems, if you are runn ing wi th more th an one LPAR for e ach CEC, m a ke sure yo u ar e r unn i n g a ff i nit y LP AR s . T o c hec k a ff i n it y b et we e n C P U, me mor y, a nd HP S lin ks , r u n t he assoc iativ i t y scri pts o n t he LPA Rs. To ch e ck the m ...
-
IBM pSeries - page 21
pshps t unin ggui d ewp0 401 05. doc P a ge 21 On the HMC GU I, selec t Se r vice A pplic ation s -> Se rvice Fo c al Poi n t - > Sele ct Se r vice a ble Even ts. 5.7 errpt command On AI X 5L, t he errpt c o mma n d list s a s u mmar y of s ys t e m er r or mes s a g es. S ome of t he H P S su bs y st e m e rro rs are co ll e c ted by e rrpt ...
-
IBM pSeries - page 22
pshps t unin ggui d ewp0 401 05. doc P a ge 22 • Fo r HA L l i b rari e s: ds h su m /u sr /s ni /a ix 52 /li b /l ib ha l_ r. a • Fo r MP I l i bra ri e s: ds h su m /u sr /l pp /p pe .p oe/ l ib /l ib mp i_ r. a (o r run with MP_P RI NTEN V=ye s) T o mak e s u r e y ou a r e r unni n g t h e c or r ect c o mb i na t i o n of H AL, L AP I, a n ...
-
IBM pSeries - page 23
pshps t unin ggui d ewp0 401 05. doc P a ge 23 MEMORY _A F FINI TY Single Thre ad Usage(MP_SINGLE_THREA D) Hin ts Fi l te red (MP_H IN TS_ FIL TE RED ) MP I-I/ O Buff er S ize (MP_IO_B UFFER _S IZE) MP I-I/O Err or Lo ggin g (MP _IO_ ER R LOG) MPI-I/ O Node File (MP_IO_ NODEF ILE) MPI-I/ O T ask List (MP _ IO_T ASKLIST ) Syst em Ch eckp o i nta b l ...
-
IBM pSeries - page 24
pshps t unin ggui d ewp0 401 05. doc P a ge 24 MPCI: se nds = 14 MPCI: se nd sComple te = 14 MPCI: se nd Wai tsC omple t e = 17 MPCI: recv s = 17 MPCI: recv WaitsCom plete = 13 M PCI : e arl yA rri v al s = 5 MPCI: e a r lyA rr iv alsMatched = 5 MP C I: la t eArr i va ls = 8 MPCI: sh ove s = 10 MP C I: p ulls = 1 3 M PCI : thre ade d L ock Y ie ld ...
-
IBM pSeries - page 25
pshps t unin ggui d ewp0 401 05. doc P a ge 25 Run the follow ing c omm a nd: /usr/sbi n /ifsn _dump - a T he dat a is coll ect ed in sn i .sn ap ( sn i _dum p .ou t .Z ) , and pro vide s usef ul in form ation , such as th e loca l mac a ddr ess: m a c_addr 0:0: 0 :40:0:0 If yo u a r e see ing arpq d rops, e n sure the source h a s the c orre ct m ...
-
IBM pSeries - page 26
pshps t unin ggui d ewp0 401 05. doc P a ge 26 To he l p you i solate the e xact cause of packe t dr ops, the i f s n_dum p -a c o mma n d a ls o lis ts t h e follow i ng debug statistic s. I f y o u i solate packe t d rops to the s e statisti cs, you wi ll probably need to con tact IB M suppo rt. dbg: | sNet_drop 0x00000000 [0] | sRTF_drop 0x00000 ...
-
IBM pSeries - page 27
pshps t unin ggui d ewp0 401 05. doc P a ge 27 T h er e a r e t wo r ou t es . sending packe t using r out e No . 1 ml ip ad d r ess structu re , sta rting : ml fl ag (ml in t erface up o r down) = 0x00000000 m l ti ck = 0 m l ip add ress = 0xc 0a80203, 19 2.168.2 .3 T h er e a r e t w o p r ef er r ed r ou t e p a ir s : f rom loc al if 0 t o re m ...
-
IBM pSeries - page 28
pshps t unin ggui d ewp0 401 05. doc P a ge 28 MA C WOF ( 2F870): B i t: 1 [. . .] 5.12.4 P ack ets d ropp ed in th e s w i tch h ardw are If a pa c k et is dr op p ed wit h i n t he s w it c h ha r d wa r e it s elf (f or ex a mp l e, wh en t r a ver s i n g th e l i nk b et w e e n t w o s w i t c h c hip s ) , e vi d e n c e o f t h e p a c k et ...
-
IBM pSeries - page 29
pshps t unin ggui d ewp0 401 05. doc P a ge 29 5.14 LAPI _DEBUG_COMM_TIMEOUT If the L API proto col e xperience s c omm unication time outs, se t the envi ronme nt v ariable LA PI_DEBUG_C OMM_T IMEOUT to PAUS E . Thi s cause s the appl ic ation to issue a pause( ) ca ll w he n enc ou nt er i n g a t ime ou t , w hi c h s t op s t h e a pp li c a t ...
-
IBM pSeries - page 30
pshps t unin ggui d ewp0 401 05. doc P a ge 30 5.16 AIX 5L trace for daemon activi ty If yo u suspect th a t a sy stem da emon is causi ng a pe r form a nce problem on yo ur sy st em , run AIX 5 L t r a ce t o c h ec k f or da em o n a c t i vit y. F or ex a mp l e, t o f i n d ou t w hic h da emo n s a r e ta kin g up CPU tim e , use the f ollowin ...
-
IBM pSeries - page 31
pshps t unin ggui d ewp0 401 05. doc P a ge 31 7.2 MPI document ation P arallel En vironme nt f or AIX 5L V4 .1.1 Hitc hhike r's G uide, SA 22- 7947-01 P arallel Env ironmen t for A IX 5L V 4.1. 1 Ope ration and Use, V olume 1 , SA22-7948- 01 P arallel Env ironmen t for A IX 5L V 4.1. 1 Ope ration and Use, V olume 2 , SA 22-7949-01 P arallel E ...
-
IBM pSeries - page 32
pshps t unin ggui d ewp0 401 05. doc P a ge 32 © IBM Cor poration 20 05 IBM Corporati on Marketing Com m unicati ons System s Gr oup Route 10 0 Somer s, New York 1 0589 Produced i n the Uni ted States of Amer ica April 2005 All Rights R eserved T his docum ent was developed for pro ducts and/or s ervices offered in the Uni ted Stat es. IBM ma y no ...
Avez-vous une question à propos de IBM pSeries?
Utiliser le formulaire ci-dessous
Si vous n’avez pas résolu votre problème avec IBM pSeries, avec l'aide du manuel que vous avez trouvé, posez une question en utilisant le formulaire ci-dessous. Si un utilisateur a eu un problème similaire avec IBM pSeries il est probable qu’il a envie de partager la façon de le résoudre.