1 User Interface for Resource Control feature
3 Intel refers to this feature as Intel Resource Director Technology(Intel(R) RDT).
4 AMD refers to this feature as AMD Platform Quality of Service(AMD QoS).
6 Copyright (C) 2016 Intel Corporation
8 Fenghua Yu <fenghua.yu@intel.com>
9 Tony Luck <tony.luck@intel.com>
10 Vikas Shivappa <vikas.shivappa@intel.com>
12 This feature is enabled by the CONFIG_X86_RESCTRL and the x86 /proc/cpuinfo
14 RDT (Resource Director Technology) Allocation - "rdt_a"
15 CAT (Cache Allocation Technology) - "cat_l3", "cat_l2"
16 CDP (Code and Data Prioritization ) - "cdp_l3", "cdp_l2"
17 CQM (Cache QoS Monitoring) - "cqm_llc", "cqm_occup_llc"
18 MBM (Memory Bandwidth Monitoring) - "cqm_mbm_total", "cqm_mbm_local"
19 MBA (Memory Bandwidth Allocation) - "mba"
21 To use the feature mount the file system:
23 # mount -t resctrl resctrl [-o cdp[,cdpl2][,mba_MBps]] /sys/fs/resctrl
27 "cdp": Enable code/data prioritization in L3 cache allocations.
28 "cdpl2": Enable code/data prioritization in L2 cache allocations.
29 "mba_MBps": Enable the MBA Software Controller(mba_sc) to specify MBA
32 L2 and L3 CDP are controlled seperately.
34 RDT features are orthogonal. A particular system may support only
35 monitoring, only control, or both monitoring and control. Cache
36 pseudo-locking is a unique way of using cache control to "pin" or
37 "lock" data in the cache. Details can be found in
38 "Cache Pseudo-Locking".
41 The mount succeeds if either of allocation or monitoring is present, but
42 only those files and directories supported by the system will be created.
43 For more details on the behavior of the interface during monitoring
44 and allocation, see the "Resource alloc and monitor groups" section.
49 The 'info' directory contains information about the enabled
50 resources. Each resource has its own subdirectory. The subdirectory
51 names reflect the resource names.
53 Each subdirectory contains the following files with respect to
56 Cache resource(L3/L2) subdirectory contains the following files
57 related to allocation:
59 "num_closids": The number of CLOSIDs which are valid for this
60 resource. The kernel uses the smallest number of
61 CLOSIDs of all enabled resources as limit.
63 "cbm_mask": The bitmask which is valid for this resource.
64 This mask is equivalent to 100%.
66 "min_cbm_bits": The minimum number of consecutive bits which
67 must be set when writing a mask.
69 "shareable_bits": Bitmask of shareable resource with other executing
70 entities (e.g. I/O). User can use this when
71 setting up exclusive cache partitions. Note that
72 some platforms support devices that have their
73 own settings for cache use which can over-ride
75 "bit_usage": Annotated capacity bitmasks showing how all
76 instances of the resource are used. The legend is:
77 "0" - Corresponding region is unused. When the system's
78 resources have been allocated and a "0" is found
79 in "bit_usage" it is a sign that resources are
81 "H" - Corresponding region is used by hardware only
82 but available for software use. If a resource
83 has bits set in "shareable_bits" but not all
84 of these bits appear in the resource groups'
85 schematas then the bits appearing in
86 "shareable_bits" but no resource group will
88 "X" - Corresponding region is available for sharing and
89 used by hardware and software. These are the
90 bits that appear in "shareable_bits" as
91 well as a resource group's allocation.
92 "S" - Corresponding region is used by software
93 and available for sharing.
94 "E" - Corresponding region is used exclusively by
95 one resource group. No sharing allowed.
96 "P" - Corresponding region is pseudo-locked. No
99 Memory bandwitdh(MB) subdirectory contains the following files
100 with respect to allocation:
102 "min_bandwidth": The minimum memory bandwidth percentage which
105 "bandwidth_gran": The granularity in which the memory bandwidth
106 percentage is allocated. The allocated
107 b/w percentage is rounded off to the next
108 control step available on the hardware. The
109 available bandwidth control steps are:
110 min_bandwidth + N * bandwidth_gran.
112 "delay_linear": Indicates if the delay scale is linear or
113 non-linear. This field is purely informational
116 If RDT monitoring is available there will be an "L3_MON" directory
117 with the following files:
119 "num_rmids": The number of RMIDs available. This is the
120 upper bound for how many "CTRL_MON" + "MON"
121 groups can be created.
123 "mon_features": Lists the monitoring events if
124 monitoring is enabled for the resource.
126 "max_threshold_occupancy":
127 Read/write file provides the largest value (in
128 bytes) at which a previously used LLC_occupancy
129 counter can be considered for re-use.
131 Finally, in the top level of the "info" directory there is a file
132 named "last_cmd_status". This is reset with every "command" issued
133 via the file system (making new directories or writing to any of the
134 control files). If the command was successful, it will read as "ok".
135 If the command failed, it will provide more information that can be
136 conveyed in the error returns from file operations. E.g.
138 # echo L3:0=f7 > schemata
139 bash: echo: write error: Invalid argument
140 # cat info/last_cmd_status
141 mask f7 has non-consecutive 1-bits
143 Resource alloc and monitor groups
144 ---------------------------------
146 Resource groups are represented as directories in the resctrl file
147 system. The default group is the root directory which, immediately
148 after mounting, owns all the tasks and cpus in the system and can make
149 full use of all resources.
151 On a system with RDT control features additional directories can be
152 created in the root directory that specify different amounts of each
153 resource (see "schemata" below). The root and these additional top level
154 directories are referred to as "CTRL_MON" groups below.
156 On a system with RDT monitoring the root directory and other top level
157 directories contain a directory named "mon_groups" in which additional
158 directories can be created to monitor subsets of tasks in the CTRL_MON
159 group that is their ancestor. These are called "MON" groups in the rest
162 Removing a directory will move all tasks and cpus owned by the group it
163 represents to the parent. Removing one of the created CTRL_MON groups
164 will automatically remove all MON groups below it.
166 All groups contain the following files:
169 Reading this file shows the list of all tasks that belong to
170 this group. Writing a task id to the file will add a task to the
171 group. If the group is a CTRL_MON group the task is removed from
172 whichever previous CTRL_MON group owned the task and also from
173 any MON group that owned the task. If the group is a MON group,
174 then the task must already belong to the CTRL_MON parent of this
175 group. The task is removed from any previous MON group.
179 Reading this file shows a bitmask of the logical CPUs owned by
180 this group. Writing a mask to this file will add and remove
181 CPUs to/from this group. As with the tasks file a hierarchy is
182 maintained where MON groups may only include CPUs owned by the
183 parent CTRL_MON group.
184 When the resouce group is in pseudo-locked mode this file will
185 only be readable, reflecting the CPUs associated with the
186 pseudo-locked region.
190 Just like "cpus", only using ranges of CPUs instead of bitmasks.
193 When control is enabled all CTRL_MON groups will also contain:
196 A list of all the resources available to this group.
197 Each resource has its own line and format - see below for details.
200 Mirrors the display of the "schemata" file to display the size in
201 bytes of each allocation instead of the bits representing the
205 The "mode" of the resource group dictates the sharing of its
206 allocations. A "shareable" resource group allows sharing of its
207 allocations while an "exclusive" resource group does not. A
208 cache pseudo-locked region is created by first writing
209 "pseudo-locksetup" to the "mode" file before writing the cache
210 pseudo-locked region's schemata to the resource group's "schemata"
211 file. On successful pseudo-locked region creation the mode will
212 automatically change to "pseudo-locked".
214 When monitoring is enabled all MON groups will also contain:
217 This contains a set of files organized by L3 domain and by
218 RDT event. E.g. on a system with two L3 domains there will
219 be subdirectories "mon_L3_00" and "mon_L3_01". Each of these
220 directories have one file per event (e.g. "llc_occupancy",
221 "mbm_total_bytes", and "mbm_local_bytes"). In a MON group these
222 files provide a read out of the current value of the event for
223 all tasks in the group. In CTRL_MON groups these files provide
224 the sum for all tasks in the CTRL_MON group and all tasks in
225 MON groups. Please see example section for more details on usage.
227 Resource allocation rules
228 -------------------------
229 When a task is running the following rules define which resources are
232 1) If the task is a member of a non-default group, then the schemata
233 for that group is used.
235 2) Else if the task belongs to the default group, but is running on a
236 CPU that is assigned to some specific group, then the schemata for the
239 3) Otherwise the schemata for the default group is used.
241 Resource monitoring rules
242 -------------------------
243 1) If a task is a member of a MON group, or non-default CTRL_MON group
244 then RDT events for the task will be reported in that group.
246 2) If a task is a member of the default CTRL_MON group, but is running
247 on a CPU that is assigned to some specific group, then the RDT events
248 for the task will be reported in that group.
250 3) Otherwise RDT events for the task will be reported in the root level
254 Notes on cache occupancy monitoring and control
255 -----------------------------------------------
256 When moving a task from one group to another you should remember that
257 this only affects *new* cache allocations by the task. E.g. you may have
258 a task in a monitor group showing 3 MB of cache occupancy. If you move
259 to a new group and immediately check the occupancy of the old and new
260 groups you will likely see that the old group is still showing 3 MB and
261 the new group zero. When the task accesses locations still in cache from
262 before the move, the h/w does not update any counters. On a busy system
263 you will likely see the occupancy in the old group go down as cache lines
264 are evicted and re-used while the occupancy in the new group rises as
265 the task accesses memory and loads into the cache are counted based on
266 membership in the new group.
268 The same applies to cache allocation control. Moving a task to a group
269 with a smaller cache partition will not evict any cache lines. The
270 process may continue to use them from the old partition.
272 Hardware uses CLOSid(Class of service ID) and an RMID(Resource monitoring ID)
273 to identify a control group and a monitoring group respectively. Each of
274 the resource groups are mapped to these IDs based on the kind of group. The
275 number of CLOSid and RMID are limited by the hardware and hence the creation of
276 a "CTRL_MON" directory may fail if we run out of either CLOSID or RMID
277 and creation of "MON" group may fail if we run out of RMIDs.
279 max_threshold_occupancy - generic concepts
280 ------------------------------------------
282 Note that an RMID once freed may not be immediately available for use as
283 the RMID is still tagged the cache lines of the previous user of RMID.
284 Hence such RMIDs are placed on limbo list and checked back if the cache
285 occupancy has gone down. If there is a time when system has a lot of
286 limbo RMIDs but which are not ready to be used, user may see an -EBUSY
289 max_threshold_occupancy is a user configurable value to determine the
290 occupancy at which an RMID can be freed.
292 Schemata files - general concepts
293 ---------------------------------
294 Each line in the file describes one resource. The line starts with
295 the name of the resource, followed by specific values to be applied
296 in each of the instances of that resource on the system.
300 On current generation systems there is one L3 cache per socket and L2
301 caches are generally just shared by the hyperthreads on a core, but this
302 isn't an architectural requirement. We could have multiple separate L3
303 caches on a socket, multiple cores could share an L2 cache. So instead
304 of using "socket" or "core" to define the set of logical cpus sharing
305 a resource we use a "Cache ID". At a given cache level this will be a
306 unique number across the whole system (but it isn't guaranteed to be a
307 contiguous sequence, there may be gaps). To find the ID for each logical
308 CPU look in /sys/devices/system/cpu/cpu*/cache/index*/id
310 Cache Bit Masks (CBM)
311 ---------------------
312 For cache resources we describe the portion of the cache that is available
313 for allocation using a bitmask. The maximum value of the mask is defined
314 by each cpu model (and may be different for different cache levels). It
315 is found using CPUID, but is also provided in the "info" directory of
316 the resctrl file system in "info/{resource}/cbm_mask". X86 hardware
317 requires that these masks have all the '1' bits in a contiguous block. So
318 0x3, 0x6 and 0xC are legal 4-bit masks with two bits set, but 0x5, 0x9
319 and 0xA are not. On a system with a 20-bit mask each bit represents 5%
320 of the capacity of the cache. You could partition the cache into four
321 equal parts with masks: 0x1f, 0x3e0, 0x7c00, 0xf8000.
323 Memory bandwidth Allocation and monitoring
324 ------------------------------------------
326 For Memory bandwidth resource, by default the user controls the resource
327 by indicating the percentage of total memory bandwidth.
329 The minimum bandwidth percentage value for each cpu model is predefined
330 and can be looked up through "info/MB/min_bandwidth". The bandwidth
331 granularity that is allocated is also dependent on the cpu model and can
332 be looked up at "info/MB/bandwidth_gran". The available bandwidth
333 control steps are: min_bw + N * bw_gran. Intermediate values are rounded
334 to the next control step available on the hardware.
336 The bandwidth throttling is a core specific mechanism on some of Intel
337 SKUs. Using a high bandwidth and a low bandwidth setting on two threads
338 sharing a core will result in both threads being throttled to use the
339 low bandwidth. The fact that Memory bandwidth allocation(MBA) is a core
340 specific mechanism where as memory bandwidth monitoring(MBM) is done at
341 the package level may lead to confusion when users try to apply control
342 via the MBA and then monitor the bandwidth to see if the controls are
343 effective. Below are such scenarios:
345 1. User may *not* see increase in actual bandwidth when percentage
346 values are increased:
348 This can occur when aggregate L2 external bandwidth is more than L3
349 external bandwidth. Consider an SKL SKU with 24 cores on a package and
350 where L2 external is 10GBps (hence aggregate L2 external bandwidth is
351 240GBps) and L3 external bandwidth is 100GBps. Now a workload with '20
352 threads, having 50% bandwidth, each consuming 5GBps' consumes the max L3
353 bandwidth of 100GBps although the percentage value specified is only 50%
354 << 100%. Hence increasing the bandwidth percentage will not yeild any
355 more bandwidth. This is because although the L2 external bandwidth still
356 has capacity, the L3 external bandwidth is fully used. Also note that
357 this would be dependent on number of cores the benchmark is run on.
359 2. Same bandwidth percentage may mean different actual bandwidth
360 depending on # of threads:
362 For the same SKU in #1, a 'single thread, with 10% bandwidth' and '4
363 thread, with 10% bandwidth' can consume upto 10GBps and 40GBps although
364 they have same percentage bandwidth of 10%. This is simply because as
365 threads start using more cores in an rdtgroup, the actual bandwidth may
366 increase or vary although user specified bandwidth percentage is same.
368 In order to mitigate this and make the interface more user friendly,
369 resctrl added support for specifying the bandwidth in MBps as well. The
370 kernel underneath would use a software feedback mechanism or a "Software
371 Controller(mba_sc)" which reads the actual bandwidth using MBM counters
372 and adjust the memowy bandwidth percentages to ensure
374 "actual bandwidth < user specified bandwidth".
376 By default, the schemata would take the bandwidth percentage values
377 where as user can switch to the "MBA software controller" mode using
378 a mount option 'mba_MBps'. The schemata format is specified in the below
381 L3 schemata file details (code and data prioritization disabled)
382 ----------------------------------------------------------------
383 With CDP disabled the L3 schemata format is:
385 L3:<cache_id0>=<cbm>;<cache_id1>=<cbm>;...
387 L3 schemata file details (CDP enabled via mount option to resctrl)
388 ------------------------------------------------------------------
389 When CDP is enabled L3 control is split into two separate resources
390 so you can specify independent masks for code and data like this:
392 L3data:<cache_id0>=<cbm>;<cache_id1>=<cbm>;...
393 L3code:<cache_id0>=<cbm>;<cache_id1>=<cbm>;...
395 L2 schemata file details
396 ------------------------
397 L2 cache does not support code and data prioritization, so the
398 schemata format is always:
400 L2:<cache_id0>=<cbm>;<cache_id1>=<cbm>;...
402 Memory bandwidth Allocation (default mode)
403 ------------------------------------------
405 Memory b/w domain is L3 cache.
407 MB:<cache_id0>=bandwidth0;<cache_id1>=bandwidth1;...
409 Memory bandwidth Allocation specified in MBps
410 ---------------------------------------------
412 Memory bandwidth domain is L3 cache.
414 MB:<cache_id0>=bw_MBps0;<cache_id1>=bw_MBps1;...
416 Reading/writing the schemata file
417 ---------------------------------
418 Reading the schemata file will show the state of all resources
419 on all domains. When writing you only need to specify those values
420 which you wish to change. E.g.
423 L3DATA:0=fffff;1=fffff;2=fffff;3=fffff
424 L3CODE:0=fffff;1=fffff;2=fffff;3=fffff
425 # echo "L3DATA:2=3c0;" > schemata
427 L3DATA:0=fffff;1=fffff;2=3c0;3=fffff
428 L3CODE:0=fffff;1=fffff;2=fffff;3=fffff
432 CAT enables a user to specify the amount of cache space that an
433 application can fill. Cache pseudo-locking builds on the fact that a
434 CPU can still read and write data pre-allocated outside its current
435 allocated area on a cache hit. With cache pseudo-locking, data can be
436 preloaded into a reserved portion of cache that no application can
437 fill, and from that point on will only serve cache hits. The cache
438 pseudo-locked memory is made accessible to user space where an
439 application can map it into its virtual address space and thus have
440 a region of memory with reduced average read latency.
442 The creation of a cache pseudo-locked region is triggered by a request
443 from the user to do so that is accompanied by a schemata of the region
444 to be pseudo-locked. The cache pseudo-locked region is created as follows:
445 - Create a CAT allocation CLOSNEW with a CBM matching the schemata
446 from the user of the cache region that will contain the pseudo-locked
447 memory. This region must not overlap with any current CAT allocation/CLOS
448 on the system and no future overlap with this cache region is allowed
449 while the pseudo-locked region exists.
450 - Create a contiguous region of memory of the same size as the cache
452 - Flush the cache, disable hardware prefetchers, disable preemption.
453 - Make CLOSNEW the active CLOS and touch the allocated memory to load
455 - Set the previous CLOS as active.
456 - At this point the closid CLOSNEW can be released - the cache
457 pseudo-locked region is protected as long as its CBM does not appear in
458 any CAT allocation. Even though the cache pseudo-locked region will from
459 this point on not appear in any CBM of any CLOS an application running with
460 any CLOS will be able to access the memory in the pseudo-locked region since
461 the region continues to serve cache hits.
462 - The contiguous region of memory loaded into the cache is exposed to
463 user-space as a character device.
465 Cache pseudo-locking increases the probability that data will remain
466 in the cache via carefully configuring the CAT feature and controlling
467 application behavior. There is no guarantee that data is placed in
468 cache. Instructions like INVD, WBINVD, CLFLUSH, etc. can still evict
469 “locked” data from cache. Power management C-states may shrink or
470 power off cache. Deeper C-states will automatically be restricted on
471 pseudo-locked region creation.
473 It is required that an application using a pseudo-locked region runs
474 with affinity to the cores (or a subset of the cores) associated
475 with the cache on which the pseudo-locked region resides. A sanity check
476 within the code will not allow an application to map pseudo-locked memory
477 unless it runs with affinity to cores associated with the cache on which the
478 pseudo-locked region resides. The sanity check is only done during the
479 initial mmap() handling, there is no enforcement afterwards and the
480 application self needs to ensure it remains affine to the correct cores.
482 Pseudo-locking is accomplished in two stages:
483 1) During the first stage the system administrator allocates a portion
484 of cache that should be dedicated to pseudo-locking. At this time an
485 equivalent portion of memory is allocated, loaded into allocated
486 cache portion, and exposed as a character device.
487 2) During the second stage a user-space application maps (mmap()) the
488 pseudo-locked memory into its address space.
490 Cache Pseudo-Locking Interface
491 ------------------------------
492 A pseudo-locked region is created using the resctrl interface as follows:
494 1) Create a new resource group by creating a new directory in /sys/fs/resctrl.
495 2) Change the new resource group's mode to "pseudo-locksetup" by writing
496 "pseudo-locksetup" to the "mode" file.
497 3) Write the schemata of the pseudo-locked region to the "schemata" file. All
498 bits within the schemata should be "unused" according to the "bit_usage"
501 On successful pseudo-locked region creation the "mode" file will contain
502 "pseudo-locked" and a new character device with the same name as the resource
503 group will exist in /dev/pseudo_lock. This character device can be mmap()'ed
504 by user space in order to obtain access to the pseudo-locked memory region.
506 An example of cache pseudo-locked region creation and usage can be found below.
508 Cache Pseudo-Locking Debugging Interface
509 ---------------------------------------
510 The pseudo-locking debugging interface is enabled by default (if
511 CONFIG_DEBUG_FS is enabled) and can be found in /sys/kernel/debug/resctrl.
513 There is no explicit way for the kernel to test if a provided memory
514 location is present in the cache. The pseudo-locking debugging interface uses
515 the tracing infrastructure to provide two ways to measure cache residency of
516 the pseudo-locked region:
517 1) Memory access latency using the pseudo_lock_mem_latency tracepoint. Data
518 from these measurements are best visualized using a hist trigger (see
519 example below). In this test the pseudo-locked region is traversed at
520 a stride of 32 bytes while hardware prefetchers and preemption
521 are disabled. This also provides a substitute visualization of cache
523 2) Cache hit and miss measurements using model specific precision counters if
524 available. Depending on the levels of cache on the system the pseudo_lock_l2
525 and pseudo_lock_l3 tracepoints are available.
527 When a pseudo-locked region is created a new debugfs directory is created for
528 it in debugfs as /sys/kernel/debug/resctrl/<newdir>. A single
529 write-only file, pseudo_lock_measure, is present in this directory. The
530 measurement of the pseudo-locked region depends on the number written to this
532 1 - writing "1" to the pseudo_lock_measure file will trigger the latency
533 measurement captured in the pseudo_lock_mem_latency tracepoint. See
535 2 - writing "2" to the pseudo_lock_measure file will trigger the L2 cache
536 residency (cache hits and misses) measurement captured in the
537 pseudo_lock_l2 tracepoint. See example below.
538 3 - writing "3" to the pseudo_lock_measure file will trigger the L3 cache
539 residency (cache hits and misses) measurement captured in the
540 pseudo_lock_l3 tracepoint.
542 All measurements are recorded with the tracing infrastructure. This requires
543 the relevant tracepoints to be enabled before the measurement is triggered.
545 Example of latency debugging interface:
546 In this example a pseudo-locked region named "newlock" was created. Here is
547 how we can measure the latency in cycles of reading from this region and
548 visualize this data with a histogram that is available if CONFIG_HIST_TRIGGERS
550 # :> /sys/kernel/debug/tracing/trace
551 # echo 'hist:keys=latency' > /sys/kernel/debug/tracing/events/resctrl/pseudo_lock_mem_latency/trigger
552 # echo 1 > /sys/kernel/debug/tracing/events/resctrl/pseudo_lock_mem_latency/enable
553 # echo 1 > /sys/kernel/debug/resctrl/newlock/pseudo_lock_measure
554 # echo 0 > /sys/kernel/debug/tracing/events/resctrl/pseudo_lock_mem_latency/enable
555 # cat /sys/kernel/debug/tracing/events/resctrl/pseudo_lock_mem_latency/hist
559 # trigger info: hist:keys=latency:vals=hitcount:sort=hitcount:size=2048 [active]
562 { latency: 456 } hitcount: 1
563 { latency: 50 } hitcount: 83
564 { latency: 36 } hitcount: 96
565 { latency: 44 } hitcount: 174
566 { latency: 48 } hitcount: 195
567 { latency: 46 } hitcount: 262
568 { latency: 42 } hitcount: 693
569 { latency: 40 } hitcount: 3204
570 { latency: 38 } hitcount: 3484
577 Example of cache hits/misses debugging:
578 In this example a pseudo-locked region named "newlock" was created on the L2
579 cache of a platform. Here is how we can obtain details of the cache hits
580 and misses using the platform's precision counters.
582 # :> /sys/kernel/debug/tracing/trace
583 # echo 1 > /sys/kernel/debug/tracing/events/resctrl/pseudo_lock_l2/enable
584 # echo 2 > /sys/kernel/debug/resctrl/newlock/pseudo_lock_measure
585 # echo 0 > /sys/kernel/debug/tracing/events/resctrl/pseudo_lock_l2/enable
586 # cat /sys/kernel/debug/tracing/trace
591 # / _----=> need-resched
592 # | / _---=> hardirq/softirq
593 # || / _--=> preempt-depth
595 # TASK-PID CPU# |||| TIMESTAMP FUNCTION
597 pseudo_lock_mea-1672 [002] .... 3132.860500: pseudo_lock_l2: hits=4097 miss=0
600 Examples for RDT allocation usage:
604 On a two socket machine (one L3 cache per socket) with just four bits
605 for cache bit masks, minimum b/w of 10% with a memory bandwidth
608 # mount -t resctrl resctrl /sys/fs/resctrl
611 # echo "L3:0=3;1=c\nMB:0=50;1=50" > /sys/fs/resctrl/p0/schemata
612 # echo "L3:0=3;1=3\nMB:0=50;1=50" > /sys/fs/resctrl/p1/schemata
614 The default resource group is unmodified, so we have access to all parts
615 of all caches (its schemata file reads "L3:0=f;1=f").
617 Tasks that are under the control of group "p0" may only allocate from the
618 "lower" 50% on cache ID 0, and the "upper" 50% of cache ID 1.
619 Tasks in group "p1" use the "lower" 50% of cache on both sockets.
621 Similarly, tasks that are under the control of group "p0" may use a
622 maximum memory b/w of 50% on socket0 and 50% on socket 1.
623 Tasks in group "p1" may also use 50% memory b/w on both sockets.
624 Note that unlike cache masks, memory b/w cannot specify whether these
625 allocations can overlap or not. The allocations specifies the maximum
626 b/w that the group may be able to use and the system admin can configure
629 If the MBA is specified in MB(megabytes) then user can enter the max b/w in MB
630 rather than the percentage values.
632 # echo "L3:0=3;1=c\nMB:0=1024;1=500" > /sys/fs/resctrl/p0/schemata
633 # echo "L3:0=3;1=3\nMB:0=1024;1=500" > /sys/fs/resctrl/p1/schemata
635 In the above example the tasks in "p1" and "p0" on socket 0 would use a max b/w
636 of 1024MB where as on socket 1 they would use 500MB.
640 Again two sockets, but this time with a more realistic 20-bit mask.
642 Two real time tasks pid=1234 running on processor 0 and pid=5678 running on
643 processor 1 on socket 0 on a 2-socket and dual core machine. To avoid noisy
644 neighbors, each of the two real-time tasks exclusively occupies one quarter
645 of L3 cache on socket 0.
647 # mount -t resctrl resctrl /sys/fs/resctrl
650 First we reset the schemata for the default group so that the "upper"
651 50% of the L3 cache on socket 0 and 50% of memory b/w cannot be used by
654 # echo "L3:0=3ff;1=fffff\nMB:0=50;1=100" > schemata
656 Next we make a resource group for our first real time task and give
657 it access to the "top" 25% of the cache on socket 0.
660 # echo "L3:0=f8000;1=fffff" > p0/schemata
662 Finally we move our first real time task into this resource group. We
663 also use taskset(1) to ensure the task always runs on a dedicated CPU
664 on socket 0. Most uses of resource groups will also constrain which
665 processors tasks run on.
667 # echo 1234 > p0/tasks
670 Ditto for the second real time task (with the remaining 25% of cache):
673 # echo "L3:0=7c00;1=fffff" > p1/schemata
674 # echo 5678 > p1/tasks
677 For the same 2 socket system with memory b/w resource and CAT L3 the
678 schemata would look like(Assume min_bandwidth 10 and bandwidth_gran is
681 For our first real time task this would request 20% memory b/w on socket
684 # echo -e "L3:0=f8000;1=fffff\nMB:0=20;1=100" > p0/schemata
686 For our second real time task this would request an other 20% memory b/w
689 # echo -e "L3:0=f8000;1=fffff\nMB:0=20;1=100" > p0/schemata
694 A single socket system which has real-time tasks running on core 4-7 and
695 non real-time workload assigned to core 0-3. The real-time tasks share text
696 and data, so a per task association is not required and due to interaction
697 with the kernel it's desired that the kernel on these cores shares L3 with
700 # mount -t resctrl resctrl /sys/fs/resctrl
703 First we reset the schemata for the default group so that the "upper"
704 50% of the L3 cache on socket 0, and 50% of memory bandwidth on socket 0
705 cannot be used by ordinary tasks:
707 # echo "L3:0=3ff\nMB:0=50" > schemata
709 Next we make a resource group for our real time cores and give it access
710 to the "top" 50% of the cache on socket 0 and 50% of memory bandwidth on
714 # echo "L3:0=ffc00\nMB:0=50" > p0/schemata
716 Finally we move core 4-7 over to the new group and make sure that the
717 kernel and the tasks running there get 50% of the cache. They should
718 also get 50% of memory bandwidth assuming that the cores 4-7 are SMT
719 siblings and only the real time threads are scheduled on the cores 4-7.
726 The resource groups in previous examples were all in the default "shareable"
727 mode allowing sharing of their cache allocations. If one resource group
728 configures a cache allocation then nothing prevents another resource group
729 to overlap with that allocation.
731 In this example a new exclusive resource group will be created on a L2 CAT
732 system with two L2 cache instances that can be configured with an 8-bit
733 capacity bitmask. The new exclusive resource group will be configured to use
734 25% of each cache instance.
736 # mount -t resctrl resctrl /sys/fs/resctrl/
739 First, we observe that the default group is configured to allocate to all L2
745 We could attempt to create the new resource group at this point, but it will
746 fail because of the overlap with the schemata of the default group:
748 # echo 'L2:0=0x3;1=0x3' > p0/schemata
751 # echo exclusive > p0/mode
752 -sh: echo: write error: Invalid argument
753 # cat info/last_cmd_status
756 To ensure that there is no overlap with another resource group the default
757 resource group's schemata has to change, making it possible for the new
758 resource group to become exclusive.
759 # echo 'L2:0=0xfc;1=0xfc' > schemata
760 # echo exclusive > p0/mode
764 p0/schemata:L2:0=03;1=03
765 p0/size:L2:0=262144;1=262144
767 A new resource group will on creation not overlap with an exclusive resource
773 p1/schemata:L2:0=fc;1=fc
774 p1/size:L2:0=786432;1=786432
776 The bit_usage will reflect how the cache is used:
777 # cat info/L2/bit_usage
778 0=SSSSSSEE;1=SSSSSSEE
780 A resource group cannot be forced to overlap with an exclusive resource group:
781 # echo 'L2:0=0x1;1=0x1' > p1/schemata
782 -sh: echo: write error: Invalid argument
783 # cat info/last_cmd_status
784 overlaps with exclusive group
786 Example of Cache Pseudo-Locking
787 -------------------------------
788 Lock portion of L2 cache from cache id 1 using CBM 0x3. Pseudo-locked
789 region is exposed at /dev/pseudo_lock/newlock that can be provided to
790 application for argument to mmap().
792 # mount -t resctrl resctrl /sys/fs/resctrl/
795 Ensure that there are bits available that can be pseudo-locked, since only
796 unused bits can be pseudo-locked the bits to be pseudo-locked needs to be
797 removed from the default resource group's schemata:
798 # cat info/L2/bit_usage
799 0=SSSSSSSS;1=SSSSSSSS
800 # echo 'L2:1=0xfc' > schemata
801 # cat info/L2/bit_usage
802 0=SSSSSSSS;1=SSSSSS00
804 Create a new resource group that will be associated with the pseudo-locked
805 region, indicate that it will be used for a pseudo-locked region, and
806 configure the requested pseudo-locked region capacity bitmask:
809 # echo pseudo-locksetup > newlock/mode
810 # echo 'L2:1=0x3' > newlock/schemata
812 On success the resource group's mode will change to pseudo-locked, the
813 bit_usage will reflect the pseudo-locked region, and the character device
814 exposing the pseudo-locked region will exist:
818 # cat info/L2/bit_usage
819 0=SSSSSSSS;1=SSSSSSPP
820 # ls -l /dev/pseudo_lock/newlock
821 crw------- 1 root root 243, 0 Apr 3 05:01 /dev/pseudo_lock/newlock
824 * Example code to access one page of pseudo-locked cache region
833 #include <sys/mman.h>
836 * It is required that the application runs with affinity to only
837 * cores associated with the pseudo-locked region. Here the cpu
838 * is hardcoded for convenience of example.
840 static int cpuid = 2;
842 int main(int argc, char *argv[])
850 page_size = sysconf(_SC_PAGESIZE);
853 CPU_SET(cpuid, &cpuset);
854 ret = sched_setaffinity(0, sizeof(cpuset), &cpuset);
856 perror("sched_setaffinity");
860 dev_fd = open("/dev/pseudo_lock/newlock", O_RDWR);
866 mapping = mmap(0, page_size, PROT_READ | PROT_WRITE, MAP_SHARED,
868 if (mapping == MAP_FAILED) {
874 /* Application interacts with pseudo-locked memory @mapping */
876 ret = munmap(mapping, page_size);
887 Locking between applications
888 ----------------------------
890 Certain operations on the resctrl filesystem, composed of read/writes
891 to/from multiple files, must be atomic.
893 As an example, the allocation of an exclusive reservation of L3 cache
896 1. Read the cbmmasks from each directory or the per-resource "bit_usage"
897 2. Find a contiguous set of bits in the global CBM bitmask that is clear
898 in any of the directory cbmmasks
899 3. Create a new directory
900 4. Set the bits found in step 2 to the new directory "schemata" file
902 If two applications attempt to allocate space concurrently then they can
903 end up allocating the same bits so the reservations are shared instead of
906 To coordinate atomic operations on the resctrlfs and to avoid the problem
907 above, the following locking procedure is recommended:
909 Locking is based on flock, which is available in libc and also as a shell
914 A) Take flock(LOCK_EX) on /sys/fs/resctrl
915 B) Read/write the directory structure.
920 A) Take flock(LOCK_SH) on /sys/fs/resctrl
921 B) If success read the directory structure.
926 # Atomically read directory structure
927 $ flock -s /sys/fs/resctrl/ find /sys/fs/resctrl
929 # Read directory contents and create new subdirectory
932 find /sys/fs/resctrl/ > output.txt
933 mask = function-of(output.txt)
934 mkdir /sys/fs/resctrl/newres/
935 echo mask > /sys/fs/resctrl/newres/schemata
937 $ flock /sys/fs/resctrl/ ./create-dir.sh
942 * Example code do take advisory locks
943 * before accessing resctrl filesystem
945 #include <sys/file.h>
948 void resctrl_take_shared_lock(int fd)
952 /* take shared lock on resctrl filesystem */
953 ret = flock(fd, LOCK_SH);
960 void resctrl_take_exclusive_lock(int fd)
964 /* release lock on resctrl filesystem */
965 ret = flock(fd, LOCK_EX);
972 void resctrl_release_lock(int fd)
976 /* take shared lock on resctrl filesystem */
977 ret = flock(fd, LOCK_UN);
988 fd = open("/sys/fs/resctrl", O_DIRECTORY);
993 resctrl_take_shared_lock(fd);
994 /* code to read directory contents */
995 resctrl_release_lock(fd);
997 resctrl_take_exclusive_lock(fd);
998 /* code to read and write directory contents */
999 resctrl_release_lock(fd);
1002 Examples for RDT Monitoring along with allocation usage:
1004 Reading monitored data
1005 ----------------------
1006 Reading an event file (for ex: mon_data/mon_L3_00/llc_occupancy) would
1007 show the current snapshot of LLC occupancy of the corresponding MON
1008 group or CTRL_MON group.
1011 Example 1 (Monitor CTRL_MON group and subset of tasks in CTRL_MON group)
1013 On a two socket machine (one L3 cache per socket) with just four bits
1016 # mount -t resctrl resctrl /sys/fs/resctrl
1017 # cd /sys/fs/resctrl
1019 # echo "L3:0=3;1=c" > /sys/fs/resctrl/p0/schemata
1020 # echo "L3:0=3;1=3" > /sys/fs/resctrl/p1/schemata
1021 # echo 5678 > p1/tasks
1022 # echo 5679 > p1/tasks
1024 The default resource group is unmodified, so we have access to all parts
1025 of all caches (its schemata file reads "L3:0=f;1=f").
1027 Tasks that are under the control of group "p0" may only allocate from the
1028 "lower" 50% on cache ID 0, and the "upper" 50% of cache ID 1.
1029 Tasks in group "p1" use the "lower" 50% of cache on both sockets.
1031 Create monitor groups and assign a subset of tasks to each monitor group.
1033 # cd /sys/fs/resctrl/p1/mon_groups
1035 # echo 5678 > m11/tasks
1036 # echo 5679 > m12/tasks
1038 fetch data (data shown in bytes)
1040 # cat m11/mon_data/mon_L3_00/llc_occupancy
1042 # cat m11/mon_data/mon_L3_01/llc_occupancy
1044 # cat m12/mon_data/mon_L3_00/llc_occupancy
1047 The parent ctrl_mon group shows the aggregated data.
1049 # cat /sys/fs/resctrl/p1/mon_data/mon_l3_00/llc_occupancy
1052 Example 2 (Monitor a task from its creation)
1054 On a two socket machine (one L3 cache per socket)
1056 # mount -t resctrl resctrl /sys/fs/resctrl
1057 # cd /sys/fs/resctrl
1060 An RMID is allocated to the group once its created and hence the <cmd>
1061 below is monitored from its creation.
1063 # echo $$ > /sys/fs/resctrl/p1/tasks
1068 # cat /sys/fs/resctrl/p1/mon_data/mon_l3_00/llc_occupancy
1071 Example 3 (Monitor without CAT support or before creating CAT groups)
1074 Assume a system like HSW has only CQM and no CAT support. In this case
1075 the resctrl will still mount but cannot create CTRL_MON directories.
1076 But user can create different MON groups within the root group thereby
1077 able to monitor all tasks including kernel threads.
1079 This can also be used to profile jobs cache size footprint before being
1080 able to allocate them to different allocation groups.
1082 # mount -t resctrl resctrl /sys/fs/resctrl
1083 # cd /sys/fs/resctrl
1084 # mkdir mon_groups/m01
1085 # mkdir mon_groups/m02
1087 # echo 3478 > /sys/fs/resctrl/mon_groups/m01/tasks
1088 # echo 2467 > /sys/fs/resctrl/mon_groups/m02/tasks
1090 Monitor the groups separately and also get per domain data. From the
1091 below its apparent that the tasks are mostly doing work on
1094 # cat /sys/fs/resctrl/mon_groups/m01/mon_L3_00/llc_occupancy
1096 # cat /sys/fs/resctrl/mon_groups/m01/mon_L3_01/llc_occupancy
1098 # cat /sys/fs/resctrl/mon_groups/m02/mon_L3_00/llc_occupancy
1100 # cat /sys/fs/resctrl/mon_groups/m02/mon_L3_01/llc_occupancy
1104 Example 4 (Monitor real time tasks)
1105 -----------------------------------
1107 A single socket system which has real time tasks running on cores 4-7
1108 and non real time tasks on other cpus. We want to monitor the cache
1109 occupancy of the real time threads on these cores.
1111 # mount -t resctrl resctrl /sys/fs/resctrl
1112 # cd /sys/fs/resctrl
1115 Move the cpus 4-7 over to p1
1118 View the llc occupancy snapshot
1120 # cat /sys/fs/resctrl/p1/mon_data/mon_L3_00/llc_occupancy