xen: modify memory ops to be NUMA-aware
PATCH 3/6 xen: modify memory ops to be NUMA-aware
Quoted 11 lines
From [1]previous post:
This patch modifies three memory operations to be NUMA-aware:
increase_reservationpopulate_physmapmemory_exchangeThese three operations request memory from the domain heap and have been modified to distribute the request across the physical cpus of the target domain evenly. This make memory local to the physical cpus within the domain available for the guest.
Measuring the overhead has shown the distribution to be costly with at the current time, no specific benefit since the best case would be providing local memory in a multi-node guest environment. As we currently don't export this virtual domain topology to Linux, it can't take advantage of the local allocations. At this time, most domains created on NUMA machines will modify their config file parameters to ensure they fit within a single NUMA node and render the distribution code useless. This patch removes the extra logic and uses domain's vcpu 0 processor as the parameter into the heap allocation function.
Now domains will use VCPU0 to pick which node to allocate memory from (using cpu_to_node mapping) and we don't pay for logic that won't be leveraged.
[1] http://lists.xensource.com/archives/html/xen-devel/2006-07/msg00544.html
Signature · 86 lines
Ryan Harper Software Engineer; Linux Technology Center IBM Corp., Austin, Tx (512) 838-9253 T/L: 678-9253 [email protected] diffstat output: memory.c | 21 ++++++++++++++------- 1 files changed, 14 insertions(+), 7 deletions(-) Signed-off-by: Ryan Harper <[email protected]> --- diff -r f362859d88c4 xen/common/memory.c --- a/xen/common/memory.c Mon Jul 31 10:30:02 2006 -0500 +++ b/xen/common/memory.c Mon Jul 31 10:34:14 2006 -0500 @@ -40,6 +40,8 @@ increase_reservation( struct page_info *page; unsigned long i; xen_pfn_t mfn; + /* use domain's first processor for locality parameter */ + unsigned int cpu = d->vcpu[0]->processor; if ( !guest_handle_is_null(extent_list) && !guest_handle_okay(extent_list, nr_extents) ) @@ -57,8 +59,8 @@ increase_reservation( return i; } - if ( unlikely((page = alloc_domheap_pages( - d, extent_order, memflags)) == NULL) ) + if ( unlikely((page = __alloc_domheap_pages( d, cpu, + extent_order, memflags )) == NULL) ) { DPRINTK("Could not allocate order=%d extent: " "id=%d memflags=%x (%ld of %d)\n", @@ -91,6 +93,8 @@ populate_physmap( unsigned long i, j; xen_pfn_t gpfn; xen_pfn_t mfn; + /* use domain's first processor for locality parameter */ + unsigned int cpu = d->vcpu[0]->processor; if ( !guest_handle_okay(extent_list, nr_extents) ) return 0; @@ -110,8 +114,8 @@ populate_physmap( if ( unlikely(__copy_from_guest_offset(&gpfn, extent_list, i, 1)) ) goto out; - if ( unlikely((page = alloc_domheap_pages( - d, extent_order, memflags)) == NULL) ) + if ( unlikely((page = __alloc_domheap_pages( d, cpu, + extent_order, memflags )) == NULL) ) { DPRINTK("Could not allocate order=%d extent: " "id=%d memflags=%x (%ld of %d)\n", @@ -293,7 +297,7 @@ memory_exchange(XEN_GUEST_HANDLE(xen_mem unsigned long in_chunk_order, out_chunk_order; xen_pfn_t gpfn, gmfn, mfn; unsigned long i, j, k; - unsigned int memflags = 0; + unsigned int memflags = 0, cpu; long rc = 0; struct domain *d; struct page_info *page; @@ -367,6 +371,9 @@ memory_exchange(XEN_GUEST_HANDLE(xen_mem } d = current->domain; + /* use domain's first processor for locality parameter */ + cpu = d->vcpu[0]->processor; + for ( i = 0; i < (exch.in.nr_extents >> in_chunk_order); i++ ) { if ( hypercall_preempt_check() ) @@ -412,8 +419,8 @@ memory_exchange(XEN_GUEST_HANDLE(xen_mem /* Allocate a chunk's worth of anonymous output pages. */ for ( j = 0; j < (1UL << out_chunk_order); j++ ) { - page = alloc_domheap_pages( - NULL, exch.out.extent_order, memflags); + page = __alloc_domheap_pages( NULL, cpu, + exch.out.extent_order, memflags); if ( unlikely(page == NULL) ) { rc = -ENOMEM;
Quoted 22 lines · Ryan Harper
Quoted 11 lines
From [1]previous post:
This patch modifies three memory operations to be NUMA-aware:
increase_reservationpopulate_physmapmemory_exchangeThese three operations request memory from the domain heap and have been modified to distribute the request across the physical cpus of the target domain evenly. This make memory local to the physical cpus within the domain available for the guest.
Measuring the overhead has shown the distribution to be costly with at the current time, no specific benefit since the best case would be providing local memory in a multi-node guest environment. As we currently don't export this virtual domain topology to Linux, it can't take advantage of the local allocations. At this time, most domains created on NUMA machines will modify their config file parameters to ensure they fit within a single NUMA node and render the distribution code useless. This patch removes the extra logic and uses domain's vcpu 0 processor as the parameter into the heap allocation function.
Quoted 6 lines
Now domains will use VCPU0 to pick which node to allocate memory from (using
cpu_to_nodemapping) and we don't pay for logic that won't be leveraged.[1] http://lists.xensource.com/archives/html/xen-devel/2006-07/msg00544.html
-no changes
Signature · 88 lines
Ryan Harper Software Engineer; Linux Technology Center IBM Corp., Austin, Tx (512) 838-9253 T/L: 678-9253 [email protected] diffstat output: memory.c | 21 ++++++++++++++------- 1 files changed, 14 insertions(+), 7 deletions(-) Signed-off-by: Ryan Harper <[email protected]> --- Make memory hypercalls NUMA-aware diff -r fa87cea10778 xen/common/memory.c --- a/xen/common/memory.c Tue Aug 15 11:38:13 2006 -0500 +++ b/xen/common/memory.c Tue Aug 15 11:40:17 2006 -0500 @@ -40,6 +40,8 @@ increase_reservation( struct page_info *page; unsigned long i; xen_pfn_t mfn; + /* use domain's first processor for locality parameter */ + unsigned int cpu = d->vcpu[0]->processor; if ( !guest_handle_is_null(extent_list) && !guest_handle_okay(extent_list, nr_extents) ) @@ -57,8 +59,8 @@ increase_reservation( return i; } - if ( unlikely((page = alloc_domheap_pages( - d, extent_order, memflags)) == NULL) ) + if ( unlikely((page = __alloc_domheap_pages( d, cpu, + extent_order, memflags )) == NULL) ) { DPRINTK("Could not allocate order=%d extent: " "id=%d memflags=%x (%ld of %d)\n", @@ -91,6 +93,8 @@ populate_physmap( unsigned long i, j; xen_pfn_t gpfn; xen_pfn_t mfn; + /* use domain's first processor for locality parameter */ + unsigned int cpu = d->vcpu[0]->processor; if ( !guest_handle_okay(extent_list, nr_extents) ) return 0; @@ -110,8 +114,8 @@ populate_physmap( if ( unlikely(__copy_from_guest_offset(&gpfn, extent_list, i, 1)) ) goto out; - if ( unlikely((page = alloc_domheap_pages( - d, extent_order, memflags)) == NULL) ) + if ( unlikely((page = __alloc_domheap_pages( d, cpu, + extent_order, memflags )) == NULL) ) { DPRINTK("Could not allocate order=%d extent: " "id=%d memflags=%x (%ld of %d)\n", @@ -293,7 +297,7 @@ memory_exchange(XEN_GUEST_HANDLE(xen_mem unsigned long in_chunk_order, out_chunk_order; xen_pfn_t gpfn, gmfn, mfn; unsigned long i, j, k; - unsigned int memflags = 0; + unsigned int memflags = 0, cpu; long rc = 0; struct domain *d; struct page_info *page; @@ -367,6 +371,9 @@ memory_exchange(XEN_GUEST_HANDLE(xen_mem } d = current->domain; + /* use domain's first processor for locality parameter */ + cpu = d->vcpu[0]->processor; + for ( i = 0; i < (exch.in.nr_extents >> in_chunk_order); i++ ) { if ( hypercall_preempt_check() ) @@ -412,8 +419,8 @@ memory_exchange(XEN_GUEST_HANDLE(xen_mem /* Allocate a chunk's worth of anonymous output pages. */ for ( j = 0; j < (1UL << out_chunk_order); j++ ) { - page = alloc_domheap_pages( - NULL, exch.out.extent_order, memflags); + page = __alloc_domheap_pages( NULL, cpu, + exch.out.extent_order, memflags); if ( unlikely(page == NULL) ) { rc = -ENOMEM;