tpcc_current (1034761), страница 15
Текст из файла (страница 15)
It d oes not necessarily im p ly loss of volatile m em ory . When the recovery m echanismrelies on the p re-failu re contents of volatile m em ory, the m eans u sed to avoid the los s of volatile m em ory (e.g., anUninterru p tible Pow er Su p p ly) m u st be inclu d ed in the system cost calcu lation. A sam p le m echanism to su rvive aninstantaneou s interru p tion in p rocessing is an u nd o/ red o log.Comment 2: In configu rations w here m ore than one instance of an op erating system can particip ate in an atom ictransaction and are connected via a p hysical m ed iu m other than an integrated bu s (e.g., bu s extend er cable, highsp eed LAN , or other connection m ethod s betw een the m u ltip le instances of the op erating system that cou ld bevu lnerable to a loss from p hysical d isru p tion), the instantaneou s interru p tion of this com m u nication is inclu d ed inthis d efinition as an item that need s to be tested .
Interru p tion of one instance of red u nd ant connections is requ ired .Comment 3: It is not the intention of this clau se to requ ire interru p tion of com m u nication to d isk tow ers or a d isksu bsystem w here red u nd ancy exists. For exam p le, log d isks can be assu m ed to p rovid e red u nd ancy for d ata d isks.3.5.3.3Failu re of all or p arts of m em ory (loss of contents).Comment: This im p lies that all or p art of m em ory has failed . This m ay be cau sed by a loss of external p ow er or thep erm anent failu re of a m em ory board .3.5.3.4Pow er Failu reComment 1: Loss of all external p ow er to the SUT for an ind efinite tim e p eriod . This m u st inclu d e at least allp ortions of the SUT that p articip ate in the d atabase p ortions of transactions.Comment 2: The p ow er failu re requ irem ent can be satisfied by p ricing su fficient UPS‟ s to gu arantee systemavailability of all com p onents that fall u nd er the p ow er failu re requ irem ent for a p eriod of at least 30 m inu tes.
Useof a UPS p rotected configu ration m u st not introd u ce new single p oints of failu re that are not p rotected by other p artsof the configu ration. The 30-m inu te requ irem ent m ay be p roven either throu gh a m easu rem ent or throu gh acalcu lation of the 30-m inu te p ow er requ irem ents (in w atts) for the p ortion of the SUT that is p rotected m u ltip lied by1.4.Comment 3: The contribu tion of the UPS in satisfying this d u rability requ irem ent d oes not need to be tested .TPC Benchm ark™C - Stand ard Sp ecification, Revision 5.11 - Page 58 of 1303.5.4D urability TestsThe intent of these tests is to d em onstrate that all transactions w hose ou tp u t m essages have been received at theterm inal or RTE have in fact been com m itted in sp ite of any single failu re from the list in Clau se 3.5.3 and that allconsistency cond itions are still m et after the d atabase is recov ered .It is requ ired that the system crash test(s) and the loss of m em ory test(s) d escribed in Clau ses 3.5.3.2 and 3.5.3.3 bep erform ed u nd er fu ll term inal load and a fu lly scaled d atabase.
The tp m C of the test ru n(s) for Clau ses 3.5.3.2 and3.5.3.3 m u st be at least 90% of the tp m C rep orted for the benchm ark.The d u rable m ed ia failu re test(s) d escribed in Clau se 3.5.3.1 m ay be p erform ed on a su bset of the SUT configu rationand d atabase. The tp m C of the test ru n for Clau se 3.5.3.1 m u st be at least 10% of the tom C rep orted for thebenchm ark.For the SUT su bset, all m u ltip le hard w are com p onents, su ch as p rocessors and d isk/ controllers in the fu ll SUTconfigu ration, m u st be rep resented by the greater of 10% of the configu ration or tw o of each of the m u ltip lehard w are com p onents.
The d atabase m u st be scaled to at least 10% of the fu lly scaled d atabase, w ith a m inim u m oftw o w arehou ses. An excep tion to the configu ration requ irem ents stated above m ay be allow ed by the TPC Au d itorin ord er to red u ce benchm ark com p lexity.
Any su ch excep tion m u st be d ocu m ented in the attestation letter from theAu d itor. Furtherm ore, the stand ard d riving m echanism m u st be u sed in this test. The test sp on sor m u st state that tothe best of their know led ge, a fu lly scaled test w ou ld also p ass all d u rability tests.For each of the failu re typ es d efined in Clau se 3.5.3, p erform the follow ing step s:1.Com p u te the su m of D_N EXT_O_ID for all row s in the DISTRICT table to d eterm ine the cu rrent cou nt ofthe total nu m ber of ord ers (cou nt1).2.Start su bm itting TPC-C transactions. The transaction rate m u st be that d escribed above and m eet all otherrequ irem ents of a rep orted m easu rem ent interval (see Clau se 5.5), exclu d ing the requ irem ent that theinterval contain at least fou r checkp oint (see Clau se 5.5.2.2).
The SUT m u st be run at this rate for at least 5m inu tes. On the Driver System , record com m itted and rolled back N ew -Ord er transaction s in a "su ccess"file.3.Cau se the failu re selected from the list in Clau se 3.5.3.4.Restart the system u nd er test u sing norm al recovery p roced u res.5.Com p are the contents of the "su ccess" file and the ORDER table to verify that every record in the "su ccess"file for a com m itted N ew -Ord er transaction has a corresp ond ing record in the ORDER table and that noentries exist for rolled back transactions.Rep eat step 1 to d eterm ine the total nu m ber of ord ers (cou nt2).
Verify that cou nt2 -cou nt1 is greater orequ al to the nu m ber of record s in the "su ccess" file for com m itted N ew -Ord er transaction s. If there is aninequ ality, the ORDER table m u st contain ad d itional record s and the d ifference m u st be less than or equ alto the nu m ber of term inals sim u lated .Comment: This d ifference shou ld be d u e only to transactions w hich w ere com m itted on the system u nd ertest, bu t for w hich the ou tp u t d ata w as not d isp layed on the inp u t/ ou tp u t screen before the failu re.6.Verify Consistency Cond ition 3 as sp ecified in Clau se 3.3.2.3.TPC Benchm ark™C - Stand ard Sp ecification, Revision 5.11 - Page 59 of 1303.5.53.5.5.1p rop erty.Additional RequirementsThe recovery m echanism cannot u se the contents of the H ISTORY table to su p p ort the d u rability3.5.5.2Roll-forw ard recovery from an archive d atabase cop y (e.g., a cop y taken p rior to the ru n) u sing red olog d ata is not accep table as the recovery m echanism in the case of failu res listed in Clau se 3.5.3.2 and 3.5.3.3.
N otethat "checkp oints", "control p oints", "consistency p oints", etc. of the d atabase taken d u ring a ru n are not con sid eredto be archives.TPC Benchm ark™C - Stand ard Sp ecification, Revision 5.11 - Page 60 of 130Clause 4: SCALIN G and D ATABASE POPULATION4.1General Scaling RulesThe throu ghp u t of the TPC-C benchm ark is d riven by the activity of the term inals connected to each w arehou se. Toincrease the throu ghp u t, m ore w arehou ses and their associated term inals m u st be configu red .
Each w arehou serequ ires a nu m ber of row s to p op u late the d atabase along w ith som e storage sp ace to m aintain the d ata generatedd u ring a d efined p eriod of activity called 60-day period. These requ irem ents d efine how storage sp ace and d atabasep op u lation scale w ith throu ghp u t.4.1.1The intent of the scaling requ irem ents is to m aintain the ratio betw een the transaction load p resentedto the system u nd er test, the card inality of the tables accessed by the transactions, the requ ired sp ace for storage, andthe nu m ber of term inals generating the transaction load .4.1.2Shou ld any scaling valu e in Clau se 4.2 be exceed ed , the others m u st be increased p rop ortionally tom aintain the sam e ratios am ong them as in Clau se 4.2.4.1.3The rep orted throu ghp u t m ay not exceed the m axim u m allow ed by the scaling requ irem ents in Clau se4.2 and the pacing requ irem ents in Clau se 5.2.
While the rep orted throu ghp u t m ay fall short of the m axim u mallow ed by the configu red system , the p rice/ p erform ance com p u tation (see Clau se 7.1) m u st rep ort the p rice for thesystem as actu ally configu red . To p revent over -scaling of system s, the rep orted throu ghp u t cannot fall short of 9tp m C p er configu red w arehou se.Comment: The m axim u m throu ghp u t is achieved w ith infinitely fast transactions resu lting in a nu ll resp onse tim eand m inim um requ ired w ait tim es.
The intent of this clau se is to p revent rep orting a throu ghp u t that exceed s thism axim u m , w hich is com p u ted to be 12.86 tp m C p er w arehou se. The above 9 tp m C rep resents 70% of the com p u tedm axim u m throu ghp u t.4.2Scaling Requirements4.2.1The WAREH OUSE table is u sed as the base u nit of scaling. The card inality of all other tables (excep tfor ITEM) is a fu nction of the nu m ber of configu red w arehou ses (i.e., card inality of the WAREH OUSE table). Thisnu m ber, in tu rn, d eterm ines the load ap p lied to the system u nd er test w hich resu lts in a rep orted throu gh p u t (seeClau se 5.4).Comment 1: The card inality of the H ISTORY, N EW-ORDER, ORDER, and ORDER-LIN E tables w ill natu rally varyas a resu lt of rep eated test execu tions. The initial d atabase p op u lation and the transaction p rofiles are d esigned tom inim ize the im p act of this variation on p erform ance and m aintain rep eatability betw een su bsequ ent test resu lts.Comment 2: The card inality of the ITEM table is constant regard less of the nu m ber of configu red w arehou ses, as allw arehou ses m aintain stocks for the sam e catalog of item s.4.2.2ConfigurationThe follow ing scaling requ irem ents rep resent the initial configu ration for the test d escribed in Clau se 5:TPC Benchm ark™C - Stand ard Sp ecification, Revision 5.11 - Page 61 of 1301.For each active w arehou se in the d atabase, the SUT m u st accep t requ ests for transactions from ap op u lation of 10 term inals.2.For each table that com p oses the d atabase, the card inality of the initial p op u lation p er w arehou se issp ecified as follow s:Table N am eWAREH OUSEDISTRICTCUSTOMERH ISTORY 1ORDER 4N EW-ORDER 4ORDER-LIN E 4STOCKITEM 2Card inality(in row s)11030k30k30k9k300k100k100kTyp ical 3 RowLength (in bytes)8995655462485430682Typ ical 3 TableSize (in 1,000 bytes)0.0890.95019,6501,3807207216,20030,6008,2001Sm all variations: su bject to test execu tion as row s m ay be inserted and d eleted by transaction ac tivityfrom test execu tions.2 Fixed card inality: d oes not scale w ith nu m ber of w arehou ses.3Typ ical lengths and sizes given here are exam p les, not requ irem ents, of w hat cou ld resu lt from anim p lem entation (sizes d o not inclu d e storage/ access overhead s).4One p ercent (1%) variation in row card inality is allow ed to accou nt for the rand om variationencou ntered d u ring the initial d ata load ing of the d atabase.N ote: The sym bol "k" u sed in the card inality colu m n m eans one thou sand3.Storage m u st be p riced for su fficient sp ace to store and m aintain the d ata generated d u ring a p eriod of 60d ays of activity w ith an average of 8 hou rs p er d ay at the rep orted th rou ghp u t called the 60-day period).This sp ace m u st be com p u ted accord ing to Clau se 4.2.3 and m u st be u sable by the d ata m anager to storeand m aintain the row s that w ou ld be ad d ed to the HISTORY, ORDER, and ORDER-LIN E tables d u ringthe 60-d ay p eriod .4.The increm ent (granu larity) for scaling the d atabase and the term inal p op u lation is one w arehou se,com p rised of one WAREH OUSE row , 10 DISTRICT row s, their associated CUSTOMER, H ISTORY,ORDER, N EW-ORDER, and ORDER-LIN E row s, 100,000 STOCK row s, 10 term inals, and p riced storagefor the 60-d ay p eriod .Comment: Over-scaling the d atabase, i.e., configu ring a larger nu m ber of w arehou ses and associated tables (Wc )than w hat is actu ally accessed d u ring the m easu rem ent (Wa) is p erm itted , p rovid ed the follow ing cond itions arem et:Let,Wc = nu m ber of w arehou ses configu red at d atabase generation,Wa = nu m ber of w arehou ses accessed d u ring the m easu rem ent (active w arehou ses),Wi = nu m ber of w arehou ses not accessed d u ring the m easu rem ent (inactive w arehou ses). It can be d em onstrated that inactive w arehou ses are not a ccessed d u ring the m easu rem ent.