No abstract

No abstract

Thas paper presents a low-overhead, hagh-performance cache coherence protocol desagned to explozt hagh bandwadth poant-to-poant and broadcast features of optacs. SPEED antegrates the vartues of snoopy-based schemes and darectory-based schemes anto one e f icaent protocol. Darectory-assast zs used exclusavely for read trafic to elamanate unnecessary broadcasts whale snoopy-assast as used exclusavely for wrate and synchronazatzon trafic to reduce darectory overhead and synchronazataon complexataes. The proposed protocol has the potentaal to ancrease performance OS Q result of ats global andependence between read and wrate operataons, concurrency an channel access, reduced contentaon, and eficaent broadcast of coherence and synchronazataon events.OMIA for memory-block-request operations (MBRs) , e.g., block transfer actions due to read-misses and write-misses. The motivation behind SPEED and OMIA's resource allocation is the following. First, COS are less frequent than MBRs [6, 71; therefore, the number of broadcasts are only a small fraction of the total number of references over the network. Second, because COS implicitly require synchronization, 0-8186-7101-7/95 $4.00 0 1995 IEEE

No abstract