¿­·¢¡¤(Öйú)ÍøÕ¾-AGÆì½¢Ìü

ÎÒÏëÏàʶ
ÓïÖÖ
ÖÐÎļòÌå ÖÐÎÄ·±Ìå English
ÓªÒµÌü
ÍøÉÏÓªÒµÌü ÕÆÉÏÓªÒµÌü
·µ»Ø¶¥²¿
ÑëÆóÊ״Σ¡Öйú¿­·¢¡¤(Öйú)ÍøÕ¾-AGÆì½¢ÌüÔÆÅÌËãÑо¿ÔºÂÛÎı»ÅÌËã»úÍøÂçÁìÓò¶¥¼¶¾Û»áÈÎÃü
2026-05-14 ÔÆÅÌËãÑо¿Ôº

¿ËÈÕ £¬£¬£¬£¬£¬£¬£¬£¬ÓÉÖйú¿­·¢¡¤(Öйú)ÍøÕ¾-AGÆì½¢ÌüÔÆÅÌËãÑо¿Ôº×÷ΪµÚÒ»µ¥Î»Íê³ÉµÄ×ÔÁ¦×ÔÖ÷Ñо¿Ð§¹û¡¶LEVELLER: Fair Communication Scheduling via Progress-Rate Awareness in Multi-Tenant Training Clusters¡·±»¹ú¼ÊÅÌËã»úÍøÂçÁìÓò¶¥¼¶¾Û»áACM SIGCOMM£¨ACM Special Interest Group on Data Communication Conference£©2026ÕýʽÈÎÃü¡£¡£¡£¡£¡£¡£¡£¡£¸ÃЧ¹ûʵÏÖÁËÑëÆóÒÔµÚÒ»µ¥Î»Éí·ÝÔÚACM SIGCOMM½ÒÏþ×ÔÁ¦×ÔÑÐÂÛÎĵÄÀúÊ·ÐÔÍ»ÆÆ £¬£¬£¬£¬£¬£¬£¬£¬±ê¼Ç×ÅÖйú¿­·¢¡¤(Öйú)ÍøÕ¾-AGÆì½¢ÌüÔÚÔÆÅÌËãÍøÂç»ù´¡Ñо¿ÓëÏÏû³Á¢Òì·½ÃæÈ¡µÃÖ÷ҪϣÍû¡£¡£¡£¡£¡£¡£¡£¡£

ACM SIGCOMMÊÇÅÌËã»úÍøÂçÁìÓò×î¾ßÓ°ÏìÁ¦µÄ¹ú¼Ê¶¥¼¶Ñ§Êõ¾Û»áÖ®Ò» £¬£¬£¬£¬£¬£¬£¬£¬±»ÖйúÅÌËã»úѧ»áÍÆ¼öĿ¼ÁÐΪCCF AÀà¾Û»á¡£¡£¡£¡£¡£¡£¡£¡£ÎåÊ®¶àÄêÀ´ £¬£¬£¬£¬£¬£¬£¬£¬SIGCOMM½ÒÏþµÄÖî¶à¾­µäÑо¿Ò»Á¬Íƶ¯Êý¾ÝͨѶϵͳ½á¹¹¡¢ÍøÂçЭÒé¡¢Êý¾ÝÖÐÐÄÍøÂçºÍ»¥ÁªÍø»ù´¡ÉèÊ©µÄÑݽø £¬£¬£¬£¬£¬£¬£¬£¬Éî¿ÌÓ°ÏìÁËÏȽøÍøÂçÊÖÒÕµÄÉú³¤Æ«Ïò¡£¡£¡£¡£¡£¡£¡£¡£SIGCOMM¶ÔÂÛÎÄÖÊÁ¿ÒªÇó¼«¸ß £¬£¬£¬£¬£¬£¬£¬£¬Ç¿µ÷»ù´¡ÐÔТ˳¡¢Ç°Õ°ÐÔÓ°ÏìÓë¼áʵµÄϵͳʵÏÖ £¬£¬£¬£¬£¬£¬£¬£¬Â¼È¡Âʺã¾Ã´¦ÓڽϵÍˮƽ £¬£¬£¬£¬£¬£¬£¬£¬½üÄêÀ´Í¨³£Ô¼Îª16%¡£¡£¡£¡£¡£¡£¡£¡£ÆäÈÎÃüÂÛÎÄÍùÍùÊܵ½Ñ§Êõ½çÓ빤ҵ½çµÄÆÕ±é¹Ø×¢ £¬£¬£¬£¬£¬£¬£¬£¬¶ÔÁ¢ÒìÊÖÒÕÂ䵨ºÍ¹¤ÒµÉú³¤¾ßÓÐÖ÷ÒªÍÆÐж¯Óᣡ£¡£¡£¡£¡£¡£¡£¾Ýͳ¼Æ £¬£¬£¬£¬£¬£¬£¬£¬×èÖ¹2025Äê £¬£¬£¬£¬£¬£¬£¬£¬ÉÐδÓÐÑëÆóÒÔµÚÒ»µ¥Î»Éí·ÝÔÚACM SIGCOMMÉϽÒÏþ×ÔÁ¦×ÔÖ÷Ñз¢Ð§¹û¡£¡£¡£¡£¡£¡£¡£¡£´Ë´ÎÖйú¿­·¢¡¤(Öйú)ÍøÕ¾-AGÆì½¢ÌüÔÆÅÌËãÑо¿ÔºÒÔµÚÒ»µ¥Î»Éí·Ý £¬£¬£¬£¬£¬£¬£¬£¬½«ÍêÈ«×ÔÁ¦×ÔÖ÷Ñз¢µÄЧ¹ûдÈëÕâÒ»¹ú¼Ê¶¥¼¶¾Û»á £¬£¬£¬£¬£¬£¬£¬£¬ÊµÏÖÁËÑëÆóÔÚÅÌËã»úÍøÂç¶¥¼¶Ñ§ÊõÎę̀ÉϵÄÀúÊ·ÐÔÍ»ÆÆ¡£¡£¡£¡£¡£¡£¡£¡£

ͼ£ºLEVELLERÉè¼Æ¼Ü¹¹£¨ÉÏ£©ÓëЧ¹û£¨Ï£©

±¾´Î±»ÊÕ¼µÄÂÛÎÄ¡¶LEVELLER: Fair Communication Scheduling via Progress-Rate Awareness in Multi-Tenant Training Clusters¡· £¬£¬£¬£¬£¬£¬£¬£¬½â¾öGPU¼¯ÈºÖжà×⻧ͨѶ¹«ÕýÐÔÄÑÌâ¡£¡£¡£¡£¡£¡£¡£¡£¸ÃÊÂÇéÓÉÖйú¿­·¢¡¤(Öйú)ÍøÕ¾-AGÆì½¢ÌüÔÆÅÌËãÑо¿Ôº×ÊÉîÖ÷ÈÎÑо¿Ô±ÀîâÙ¡¢ÊµÏ°ÉúÀîã󣨱±¾©Óʵç´óѧÔÚ¶Á²©Ê¿Éú£©¡¢Ñо¿Ô±ê°Ã÷Ô¶ÒÔ¼°Öйú¿­·¢¡¤(Öйú)ÍøÕ¾-AGÆì½¢Ìü¼¯ÍÅÊ×ϯ¿ÆÑ§¼Ò¡¢Öйú¿­·¢¡¤(Öйú)ÍøÕ¾-AGÆì½¢ÌüÔÆÅÌËãÑо¿ÔºÔº³¤Îâ½Ü½ÌÊÚÍê³É¡£¡£¡£¡£¡£¡£¡£¡£

Ä¿½ñ £¬£¬£¬£¬£¬£¬£¬£¬ÃæÏò¶à×â»§µÄGPU¼¯ÈºÒѳÉΪ´óÄ£×ÓѵÁ·£¨LLM£©µÄ½¹µã»ù´¡ÉèÊ©¡£¡£¡£¡£¡£¡£¡£¡£ÔÚ¶à¸öѵÁ·Ê¹Ãü¹²ÏíÍøÂç×ÊÔ´µÄÇéÐÎÏ £¬£¬£¬£¬£¬£¬£¬£¬ÓÉÓÚʹÃüÌØÕ÷±£´æ²î±ð £¬£¬£¬£¬£¬£¬£¬£¬ÏÖÓÐÖ÷Á÷ͨѶµ÷ÀíϵͳÄÑÒÔ°ü¹Ü¹«ÕýÐÔ £¬£¬£¬£¬£¬£¬£¬£¬³£µ¼Ö²¿·ÖʹÃü±»¡°¶öËÀ¡±»ò½ø¶ÈÖͺ󡣡£¡£¡£¡£¡£¡£¡£Í¨Ñ¶²»¹«Õý £¬£¬£¬£¬£¬£¬£¬£¬²»µ«Ó°Ïì¶à×â»§µÄÓû§ÌåÑé £¬£¬£¬£¬£¬£¬£¬£¬ÖÆÔ¼¼¯ÈºµÄÕûÌåЧÄÜ £¬£¬£¬£¬£¬£¬£¬£¬¸üÖ±½ÓÍþвµ½ÖÇËãÔÆÐ§ÀͼòÖ±¶¨ÐÔ£¨Cloud Integrity£©ºÍÉÌÒµ×óȯ¡£¡£¡£¡£¡£¡£¡£¡£

Ϊ´Ë £¬£¬£¬£¬£¬£¬£¬£¬¸ÃÊÂÇéÁ¢ÒìÐÔÌá³ö¹éÒ»»¯½ø¶ÈÂÊ£¨Normalized Progress Rate£©Ö¸±ê £¬£¬£¬£¬£¬£¬£¬£¬Í¨¹ýȨºâʹÃüÔÚ¾ºÕùÇéÐÎϵÄÏÖʵ½ø¶ÈÓëÎÞ×ÌÈÅÀíÏë½ø¶ÈµÄ±ÈÀý £¬£¬£¬£¬£¬£¬£¬£¬¾«×¼Á¿»¯ÑµÁ·ÌåÑé¡£¡£¡£¡£¡£¡£¡£¡£ÕâһʹÃüÎ޹صÄÖ¸±êÀÖ³ÉÌî²¹Á˵ײãflow-level¹«ÕýÐÔÓëÉϲãÄ£×ÓѵÁ·job-level¹«ÕýÐÔÖ®¼äµÄÀíÂÛ¿Õȱ £¬£¬£¬£¬£¬£¬£¬£¬ÊǸÃÁìÓòµÄÖ÷ÒªÊÖÒÕÍ»ÆÆ £¬£¬£¬£¬£¬£¬£¬£¬»ò³ÉΪδÀ´ÐÐÒµ±ê×¼¡£¡£¡£¡£¡£¡£¡£¡£» £»£» £»£»£»£»£»ùÓÚ¸ÃÖ¸±ê £¬£¬£¬£¬£¬£¬£¬£¬Ñо¿ÍŶӹ¹½¨ÁËÍêÕûµÄ¹«ÕýÐÔÀíÂÛ £¬£¬£¬£¬£¬£¬£¬£¬²¢¿ª·¢ÁËLEVELLERϵͳ £¬£¬£¬£¬£¬£¬£¬£¬Ê×´ÎÔÚ¶à×â»§¼¯ÈºÖÐÕë¶Ôí§ÒâÊÂÇé¸ºÔØÊµÏÖͨѶµ÷ÀíµÄ×î´ó»¯-×îС»¯¹«Õý£¨Max-Min Fairness£©¡£¡£¡£¡£¡£¡£¡£¡£

LEVELLERϵͳ¼«¾ßÊÊÓÃÐÔÓë¿ÉÀ©Õ¹ÐÔ £¬£¬£¬£¬£¬£¬£¬£¬Ö§³ÖÔÚRDMAºÍTCPÏÖÓÐÓ²¼þÉÏÖ±½Ó°²ÅÅ¡£¡£¡£¡£¡£¡£¡£¡£ÊµÑéЧ¹ûÏÔʾ £¬£¬£¬£¬£¬£¬£¬£¬ÔÚ10ÖÖ´óÓïÑÔÄ£×ӵIJâÊÔÖÐ £¬£¬£¬£¬£¬£¬£¬£¬LEVELLERÏà±ÈÐÐÒµÖ÷Á÷¼Æ»® £¬£¬£¬£¬£¬£¬£¬£¬ÌáÉý×îµÍ½ø¶ÈÂÊ37% £¬£¬£¬£¬£¬£¬£¬£¬ÓÅ»¯¹«ÕýÐÔ17% £¬£¬£¬£¬£¬£¬£¬£¬Í¬Ê±¼á³Ö¼«¸ßµÄ¼¯Èº×ÊԴʹÓÃÂÊ¡£¡£¡£¡£¡£¡£¡£¡£¸ÃÊÂÇéΪ¶à×â»§AI¼¯ÈºÌṩÁËÐµĹ«ÕýÐÔ»ù×¼ £¬£¬£¬£¬£¬£¬£¬£¬Ò²ÎªÖÇËãÖÐÐÄ£¨AIDC£©´ó¹æÄ£ÑµÁ·Í¨Ñ¶µ÷ÀíÌṩÇÐʵ¿ÉÐеĽâ¾ö¼Æ»®¡£¡£¡£¡£¡£¡£¡£¡£

±ðµÄ £¬£¬£¬£¬£¬£¬£¬£¬Öйú¿­·¢¡¤(Öйú)ÍøÕ¾-AGÆì½¢ÌüÔÆÅÌËãÑо¿Ôº³Â×ÓÐùÑо¿Ô±¼ÓÈëÍê³ÉµÄÏàÖúЧ¹û¡¶Scale-up PIFO: Interleaving Multiple Priority Queues for High Speed Programmable Scheduling¡·Ò²±»ACM SIGCOMM 2026ÈÎÃü¡£¡£¡£¡£¡£¡£¡£¡£¸ÃÊÂÇéÓɸ´µ©´óѧÐìÑï½ÌÊÚ¿ÎÌâ×éǣͷ £¬£¬£¬£¬£¬£¬£¬£¬ÃæÏòAIÊý¾ÝÖÐÐĺÍÐÂÐÍÔÆÍø»ù´¡ÉèÊ©Öн»Á÷»ú¶Ë¿ÚËÙÂÊÒ»Á¬ÌáÉý´øÀ´µÄ¸ßÐÔÄܵ÷ÀíÐèÇó £¬£¬£¬£¬£¬£¬£¬£¬Õë¶Ô¹Å°åµ¥PIFOÐÐÁÐÄÑÒÔÖ§³Ö1.6Tbps¼¶ÏßËÙ´¦Öóͷ£¡¢¼òÆÓ²¢Ðл¯ÓÖ»áÒýÈëµ÷ÀíÎó²îµÈÎÊÌâ £¬£¬£¬£¬£¬£¬£¬£¬Ìá³ö¸ßËٿɱà³Ìµ÷Àí¿ò¼ÜScale-up PIFO¡£¡£¡£¡£¡£¡£¡£¡£¸Ã¿ò¼Üͨ¹ý½»Ö¯²¢Ðжà¸öPIFOÐÐÁÐÌáÉýµ÷ÀíÍÌÍ £¬£¬£¬£¬£¬£¬£¬£¬²¢Éè¼ÆRank Range Load BalancingËã·¨ £¬£¬£¬£¬£¬£¬£¬£¬ÔÚ¿ØÖƵ÷ÀíÎó²îµÄͬʱ¼á³ÖÓ²¼þʵÏֵľ«Á·ÐÔ £¬£¬£¬£¬£¬£¬£¬£¬ÎªÏÂÒ»´ú¸ßËÙÊý¾ÝÖÐÐÄÍøÂçÖеĿɱà³ÌQoSµ÷ÀíÌṩÁËеÄÊÖÒÕ·¾¶¡£¡£¡£¡£¡£¡£¡£¡£

½üÄêÀ´ £¬£¬£¬£¬£¬£¬£¬£¬ÑëÆóÔÚ»ù´¡Ñо¿ÓëԭʼÁ¢ÒìÁìÓòÒ»Á¬»ýÀÛ¡¢ºñ»ý±¡·¢¡£¡£¡£¡£¡£¡£¡£¡£Öйú¿­·¢¡¤(Öйú)ÍøÕ¾-AGÆì½¢ÌüÔÆÅÌËãÑо¿ÔºÑо¿ÍŶÓÔÚÖйú¿­·¢¡¤(Öйú)ÍøÕ¾-AGÆì½¢Ìü¼¯ÍÅÊ×ϯ¿ÆÑ§¼Ò¡¢ÔÆÅÌËãÑо¿ÔºÔº³¤Îâ½Ü½ÌÊÚµÄÏòµ¼Ï £¬£¬£¬£¬£¬£¬£¬£¬Ò»Á¬Éî¸ûÔÆÅÌËãÍøÂç»ù´¡ÊÖÒÕÓëÒªº¦ÏÏû³Á¢Òì £¬£¬£¬£¬£¬£¬£¬£¬´ÓÀíÂÛÌáÁ¶µ½ÏµÍ³ÊµÖ¤ £¬£¬£¬£¬£¬£¬£¬£¬ÔÚÃæÏòÖÇÄÜÅÌËã»ù´¡ÉèÊ©µÄÍøÂçÒªº¦ÎÊÌâÉÏ¿ªÕ¹ºã¾Ã¹¥¹Ø¡£¡£¡£¡£¡£¡£¡£¡£´Ë´ÎЧ¹ûÈÎÃüACM SIGCOMM 2026 £¬£¬£¬£¬£¬£¬£¬£¬ÌåÏÖÁËÖйú¿­·¢¡¤(Öйú)ÍøÕ¾-AGÆì½¢ÌüÔÆÅÌËãÑо¿ÔºÔÚ¹ú¼ÊÅÌËã»úÍøÂçÑо¿Ç°ÑصÄÔ­´´Á¢ÒìÄÜÁ¦ £¬£¬£¬£¬£¬£¬£¬£¬Ò²Åú×¢ÑëÆó²»µ«Äܹ»ÔÚÖØ´ó¹¤³Ì½¨ÉèÖС°¿¸´óÁº¡± £¬£¬£¬£¬£¬£¬£¬£¬ÕýÔÚ»ù´¡Ñо¿ÓëԭʼÁ¢ÒìÖÐÒ»Á¬·¢³öÖйúÆóÒµµÄÊÖÒÕÉùÒô¡£¡£¡£¡£¡£¡£¡£¡£

δÀ´ £¬£¬£¬£¬£¬£¬£¬£¬Öйú¿­·¢¡¤(Öйú)ÍøÕ¾-AGÆì½¢ÌüÔÆÅÌËãÑо¿Ôº½«¼ÌÐø¼á³ÖÁ¢ÒìÇý¶¯Éú³¤ £¬£¬£¬£¬£¬£¬£¬£¬Éî»¯ÔÆÅÌËãÍøÂç»ù´¡ÊÖÒսṹ £¬£¬£¬£¬£¬£¬£¬£¬Íƶ¯ÖØµã¿ÆÑÐЧ¹ûÏò½¹µãÊÖÒÕÄÜÁ¦×ª»¯ £¬£¬£¬£¬£¬£¬£¬£¬²¢ÈÚÈëÌìÒíÔÆÆ½Ì¨ÄÜÁ¦ÏµÍ³ £¬£¬£¬£¬£¬£¬£¬£¬Ò»Á¬ÔöǿҪº¦µ××ùÄÜÁ¦ £¬£¬£¬£¬£¬£¬£¬£¬Ò»Ö±ÌáÉý×ÔÖ÷Á¢ÒìˮƽÓëϵͳ»¯¾ºÕùÓÅÊÆ¡£¡£¡£¡£¡£¡£¡£¡£Í¬Ê± £¬£¬£¬£¬£¬£¬£¬£¬Öйú¿­·¢¡¤(Öйú)ÍøÕ¾-AGÆì½¢ÌüÔÆÅÌËãÑо¿Ôº½«½øÒ»²½Ê©Õ¹ÔÚÔÆÅÌËã¡¢ÍøÂçϵͳºÍÖÇÄÜÅÌËã»ù´¡ÉèÊ©ÁìÓòµÄÊÖÒÕ»ýÀÛÓëÈ˲ÅÓÅÊÆ £¬£¬£¬£¬£¬£¬£¬£¬ÎªÖйú¿­·¢¡¤(Öйú)ÍøÕ¾-AGÆì½¢Ìü¡°ÔÆ¡ªÍø¡ªÊý¡ªÖÇ¡±ÈÚºÏÉú³¤ÌṩԽ·¢¼áʵµÄµ×²ãÊÖÒÕÖ§³Ö¡£¡£¡£¡£¡£¡£¡£¡£

ɨһɨÔÚÊÖ»ú·­¿ªÄ¿½ñÒ³