An Evaluation of an Atte m pt at Offloading TCP/IP Protocol Proce ssing onto an i9 60RN-based iNIC

TCP/IP n e tw ork ing, Inte llige n t Ne tw ork Inte rface Th is re port pre s e nts an e valuation of a TCP/IP offload im ple m e n tation th at utiliz e s a 100Base T inte llige n t Ne tw ork Inte rface Card (iNIC) equipped w ith a 100 M H z i9 60RN processor. Th e e n tire Fre e BSD-de rived n etw ork ing stack from sock e t dow nw ard is im ple m e n ted on th e iNIC w ith th e goal of reducing h ost processor w ork load. For large m e ssages th at result in MTU pack e ts, th e offload im ple m e n tation can sustain w ire -spe e d on re ce ive but only about 80% of w ire -spe e d on transm it. Utiliz ing h ardw are -based profiling of TTCP b ench m ark runs, our e valuation pie ced toge th e r a com pre h ensive picture of transm it b e h avior on th e iNIC. Our first surprise w as th e num b e r of i9 60RN processor cycles consum e d in transm itting large m e ssages --around 17 th ousand processor cycles pe r 1.5k byte (Eth e rn e t M TU) pack e t. Furth e r investigation re ve als th at th is h igh cost is due to a com bination of i9 60RN arch ite ctural sh ortcom ings, poor buffe ring strate gy in th e TCP/IP code running on th e iNIC, and lim itations im posed by th e I20-based h ost-iNIC inte rface . W e also found room for im prove m e n ts in th e im ple m e n tation of th e sock e t buffe r data-structure . Th is re port pre s e nts profiling statistics, as w e ll as code -path analysis th at back up th e s e conclusions. Our results call into question th e h ypoth esis th at a spe cializ e d n etw ork ing softw are e n vironm e nt coupled w ith ch e ap e m bedded processors is a cost e ffe ctive w ay of im proving syste m pe rform ance . At le ast in th e case of th e offload im ple m e n tation on th e i9 60RN-based iNIC, n e ith e r w as th e pe rform ance adequate n or th e syste m ch e ap. Th is conclusion, h ow e ve r, doe s not im ply th at offload is a bad ide a. In fact, m e asure m e n ts w e m ade w ith Alacrite ch 's SLIC NIC, w h ich partially offloads TCP/IP protocol processing to an ASIC, suggests th at offloading can confe r advantages in a cost e ffe ctive w ay. Tak ing th e righ t im ple m e n tation approach is critical.