Attachment #463 for ticket #466

View | Details | Raw Unified | Return to ticket 466
Collapse All | Expand All

Lines 30-35 documents those changes that are of interest to users and admins. Link Here

(-)a/NEWS (+3 lines)
30	-- Replace the tempname() function call with mkstemp().	30	-- Replace the tempname() function call with mkstemp().
31	-- Fix for --cpu_bind=map_cpu/mask_cpu/map_ldom/mask_ldom plus	31	-- Fix for --cpu_bind=map_cpu/mask_cpu/map_ldom/mask_ldom plus
32	--mem_bind=map_mem/mask_mem options, broken in 2.6.2.	32	--mem_bind=map_mem/mask_mem options, broken in 2.6.2.
		33	-- Restore default behavior of allocating cores to jobs on a cyclic basis
		34	across the sockets unless SelectTypeParameters=CR_CORE_DEFAULT_DIST_BLOCK
		35	or user specifies other distribution options.
33		36
34	* Changes in Slurm 2.6.3	37	* Changes in Slurm 2.6.3
35	========================	38	========================




	bool *sock_used, *sock_avoid;
	bool alloc_cores = false, alloc_sockets = false;
	uint16_t ntasks_per_core = 0xffff, ntasks_per_socket = 0xffff;
	int error_code = SLURM_SUCCESS;
	uint32_t total_cpus, *cpus_cnt;

	if ((job_res == NULL) || (job_res->core_bitmap == NULL))
		return error_code;

		core_cnt = 0;
		cpus  = job_res->cpus[i];

		if (ntasks_per_socket != 0xffff) {
			int x_cpus;
			uint32_t total_cpus = 0;
			uint32_t *cpus_cnt = xmalloc(sizeof(uint32_t)* sockets);
		for (s = 0; s < sockets; s++) {
			for (j = sock_start[s]; j < sock_end[s]; j++) {
				if (bit_test(core_map, j))
					cpus_cnt[s] += vpus;
			}
			total_cpus += cpus_cnt[s];
		}
		for (s = 0; s < sockets && total_cpus > cpus; s++) {
			if ((ntasks_per_socket != 0xffff) &&
			    (cpus_cnt[s] > ntasks_per_socket)) {
				int x_cpus = cpus_cnt[s] - ntasks_per_socket;
				x_cpus = MIN(x_cpus, (total_cpus - cpus));
				cpus_cnt[s] -= x_cpus;
				total_cpus  -= x_cpus;
			}
			if ((cpus_cnt[s] >= cpus) &&
			    ((socket_best_fit == -1) ||
			     (cpus_cnt[s] < cpus_cnt[socket_best_fit])))
				socket_best_fit = s;
		}
		if (socket_best_fit != -1) {
			/* Use one socket with best fit, avoid all others */
			for (s = 0; s < sockets; s++) {
				for (j = sock_start[s]; j < sock_end[s]; j++) {
					if (bit_test(core_map, j))
						cpus_cnt[s] += vpus;
				}
				total_cpus += cpus_cnt[s];
			}
			for (s = 0; s < sockets && total_cpus > cpus; s++) {
				if (cpus_cnt[s] > ntasks_per_socket) {
					x_cpus = cpus_cnt[s] -ntasks_per_socket;
					cpus_cnt[s] = ntasks_per_socket;
					total_cpus -= x_cpus;
				}
			}
			total_cpus = cpus;
		} else if (ntasks_per_socket != 0xffff) {
			/* Avoid sockets that can't start ntasks */
			for (s = 0; s < sockets && total_cpus > cpus; s++) {
				if ((cpus_cnt[s] <= ntasks_per_socket) &&
				    (total_cpus - cpus_cnt[s] >= cpus)) {
					sock_avoid[s] = true;
					total_cpus -= cpus_cnt[s];
				}
			}
			xfree(cpus_cnt);
		}
		xfree(cpus_cnt);

		while (cpus > 0) {
			uint16_t prev_cpus = cpus;
- 

Return to ticket 466

Lines 675-682 static int _cyclic_sync_core_bitmap(struct job_record *job_ptr, Link Here

(-)a/src/plugins/select/cons_res/dist_tasks.c (-36 / +19 lines)
675	bool sock_used, sock_avoid;	675	bool sock_used, sock_avoid;
676	bool alloc_cores = false, alloc_sockets = false;	676	bool alloc_cores = false, alloc_sockets = false;
677	uint16_t ntasks_per_core = 0xffff, ntasks_per_socket = 0xffff;	677	uint16_t ntasks_per_core = 0xffff, ntasks_per_socket = 0xffff;
678	int error_code = SLURM_SUCCESS, socket_best_fit;	678	int error_code = SLURM_SUCCESS;
679	uint32_t total_cpus, *cpus_cnt;
680		679
681	if ((job_res == NULL) \|\| (job_res->core_bitmap == NULL))	680	if ((job_res == NULL) \|\| (job_res->core_bitmap == NULL))
682	return error_code;	681	return error_code;
Lines 749-796 static int _cyclic_sync_core_bitmap(struct job_record *job_ptr, Link Here
749	core_cnt = 0;	748	core_cnt = 0;
750	cpus = job_res->cpus[i];	749	cpus = job_res->cpus[i];
751		750
752	/* Pack job onto socket(s) with best fit */	751	if (ntasks_per_socket != 0xffff) {
753	socket_best_fit = -1;	752	int x_cpus;
754	total_cpus = 0;	753	uint32_t total_cpus = 0;
755	cpus_cnt = xmalloc(sizeof(uint32_t)* sockets);	754	uint32_t cpus_cnt = xmalloc(sizeof(uint32_t) sockets);
756	for (s = 0; s < sockets; s++) {
757	for (j = sock_start[s]; j < sock_end[s]; j++) {
758	if (bit_test(core_map, j))
759	cpus_cnt[s] += vpus;
760	}
761	total_cpus += cpus_cnt[s];
762	}
763	for (s = 0; s < sockets && total_cpus > cpus; s++) {
764	if ((ntasks_per_socket != 0xffff) &&
765	(cpus_cnt[s] > ntasks_per_socket)) {
766	int x_cpus = cpus_cnt[s] - ntasks_per_socket;
767	x_cpus = MIN(x_cpus, (total_cpus - cpus));
768	cpus_cnt[s] -= x_cpus;
769	total_cpus -= x_cpus;
770	}
771	if ((cpus_cnt[s] >= cpus) &&
772	((socket_best_fit == -1) \|\|
773	(cpus_cnt[s] < cpus_cnt[socket_best_fit])))
774	socket_best_fit = s;
775	}
776	if (socket_best_fit != -1) {
777	/* Use one socket with best fit, avoid all others */
778	for (s = 0; s < sockets; s++) {	755	for (s = 0; s < sockets; s++) {
779	if (s != socket_best_fit)	756	for (j = sock_start[s]; j < sock_end[s]; j++) {
780	sock_avoid[s] = true;	757	if (bit_test(core_map, j))
		758	cpus_cnt[s] += vpus;
		759	}
		760	total_cpus += cpus_cnt[s];
		761	}
		762	for (s = 0; s < sockets && total_cpus > cpus; s++) {
		763	if (cpus_cnt[s] > ntasks_per_socket) {
		764	x_cpus = cpus_cnt[s] -ntasks_per_socket;
		765	cpus_cnt[s] = ntasks_per_socket;
		766	total_cpus -= x_cpus;
		767	}
781	}	768	}
782	total_cpus = cpus;
783	} else if (ntasks_per_socket != 0xffff) {
784	/* Avoid sockets that can't start ntasks */
785	for (s = 0; s < sockets && total_cpus > cpus; s++) {	769	for (s = 0; s < sockets && total_cpus > cpus; s++) {
786	if ((cpus_cnt[s] <= ntasks_per_socket) &&	770	if ((cpus_cnt[s] <= ntasks_per_socket) &&
787	((total_cpus - cpus_cnt[s]) >= cpus)) {	771	(total_cpus - cpus_cnt[s] >= cpus)) {
788	sock_avoid[s] = true;	772	sock_avoid[s] = true;
789	total_cpus -= cpus_cnt[s];	773	total_cpus -= cpus_cnt[s];
790	}	774	}
791	}	775	}
		776	xfree(cpus_cnt);
792	}	777	}
793	xfree(cpus_cnt);
794		778
795	while (cpus > 0) {	779	while (cpus > 0) {
796	uint16_t prev_cpus = cpus;	780	uint16_t prev_cpus = cpus;
797	-