Characterizing vulnerability of parallelism to resource constraints

T h e theoret ical available i n s t r u c t i o n level paral le l i sm in m o s t b e n c h m a r k s is v e r y high. Vulnerabi l i ty i s related t o t h e di f f icul ty with w h i c h w e can extract t h i s paral le l i sm with f i n i t e resources . This s t u d y characterizes t h e vu lnerabi l i ty of paral le l i sm t o resource cons tra in ts by scheduling d y n a m i c dependence graphs ( D D G s ) f r o m traces of several b e n c h m a r k s using d i g e r e n t scheduling a lgor i thms and di f ferent n u m b e r of f u n c t i o n a l un i t s . It is observed t h a t t h e e z e c u t i o n tame of t h e DDGs does n o t v a r y s igni f icant ly w i t h low-level schedul ing algori thms l ike lazy, slack, e tc . M e a s u r e s of vulnerabi l i ty based o n slack a n d load were also considered. A l t h o u g h Accslk-Load, which u s e s a combinat i o n of accurate s lack and load t o m a k e a predict ion, has a predic t ion accuracy of about 85%, t h e predic t ion rate i s only 42%. On t h e o t h e r hand, e v e n t h o u g h t h e predic t ion accuracy of a(L,), t h e s tandard dev ia t ion in t h e load, is n o t a s h igh , t h e r e is a predic t ion in all t h e cases. T h e DDG e z e c u t i o n t i m e i s also f o u n d t o be m o s t vulnerable t o t h e f u n c t i o n a l unit with t h e greatest u(Lz).