User Interface for Resource Allocation in Intel Resource Director Technology Copyright (C) 2016 Intel Corporation Fenghua Yu Tony Luck Vikas Shivappa This feature is enabled by the CONFIG_INTEL_RDT Kconfig and the X86 /proc/cpuinfo flag bits "rdt", "cqm", "cat_l3" and "cdp_l3". To use the feature mount the file system: # mount -t resctrl resctrl [-o cdp] /sys/fs/resctrl mount options are: "cdp": Enable code/data prioritization in L3 cache allocations. RDT features are orthogonal. A particular system may support only monitoring, only control, or both monitoring and control. The mount succeeds if either of allocation or monitoring is present, but only those files and directories supported by the system will be created. For more details on the behavior of the interface during monitoring and allocation, see the "Resource alloc and monitor groups" section. Info directory -------------- The 'info' directory contains information about the enabled resources. Each resource has its own subdirectory. The subdirectory names reflect the resource names. Each subdirectory contains the following files with respect to allocation: Cache resource(L3/L2) subdirectory contains the following files related to allocation: "num_closids": The number of CLOSIDs which are valid for this resource. The kernel uses the smallest number of CLOSIDs of all enabled resources as limit. "cbm_mask": The bitmask which is valid for this resource. This mask is equivalent to 100%. "min_cbm_bits": The minimum number of consecutive bits which must be set when writing a mask. "shareable_bits": Bitmask of shareable resource with other executing entities (e.g. I/O). User can use this when setting up exclusive cache partitions. Note that some platforms support devices that have their own settings for cache use which can over-ride these bits. Memory bandwitdh(MB) subdirectory contains the following files with respect to allocation: "min_bandwidth": The minimum memory bandwidth percentage which user can request. "bandwidth_gran": The granularity in which the memory bandwidth percentage is allocated. The allocated b/w percentage is rounded off to the next control step available on the hardware. The available bandwidth control steps are: min_bandwidth + N * bandwidth_gran. "delay_linear": Indicates if the delay scale is linear or non-linear. This field is purely informational only. If RDT monitoring is available there will be an "L3_MON" directory with the following files: "num_rmids": The number of RMIDs available. This is the upper bound for how many "CTRL_MON" + "MON" groups can be created. "mon_features": Lists the monitoring events if monitoring is enabled for the resource. "max_threshold_occupancy": Read/write file provides the largest value (in bytes) at which a previously used LLC_occupancy counter can be considered for re-use. Finally, in the top level of the "info" directory there is a file named "last_cmd_status". This is reset with every "command" issued via the file system (making new directories or writing to any of the control files). If the command was successful, it will read as "ok". If the command failed, it will provide more information that can be conveyed in the error returns from file operations. E.g. # echo L3:0=f7 > schemata bash: echo: write error: Invalid argument # cat info/last_cmd_status mask f7 has non-consecutive 1-bits Resource alloc and monitor groups --------------------------------- Resource groups are represented as directories in the resctrl file system. The default group is the root directory which, immediately after mounting, owns all the tasks and cpus in the system and can make full use of all resources. On a system with RDT control features additional directories can be created in the root directory that specify different amounts of each resource (see "schemata" below). The root and these additional top level directories are referred to as "CTRL_MON" groups below. On a system with RDT monitoring the root directory and other top level directories contain a directory named "mon_groups" in which additional directories can be created to monitor subsets of tasks in the CTRL_MON group that is their ancestor. These are called "MON" groups in the rest of this document. Removing a directory will move all tasks and cpus owned by the group it represents to the parent. Removing one of the created CTRL_MON groups will automatically remove all MON groups below it. All groups contain the following files: "tasks": Reading this file shows the list of all tasks that belong to this group. Writing a task id to the file will add a task to the group. If the group is a CTRL_MON group the task is removed from whichever previous CTRL_MON group owned the task and also from any MON group that owned the task. If the group is a MON group, then the task must already belong to the CTRL_MON parent of this group. The task is removed from any previous MON group. "cpus": Reading this file shows a bitmask of the logical CPUs owned by this group. Writing a mask to this file will add and remove CPUs to/from this group. As with the tasks file a hierarchy is maintained where MON groups may only include CPUs owned by the parent CTRL_MON group. "cpus_list": Just like "cpus", only using ranges of CPUs instead of bitmasks. When control is enabled all CTRL_MON groups will also contain: "schemata": A list of all the resources available to this group. Each resource has its own line and format - see below for details. When monitoring is enabled all MON groups will also contain: "mon_data": This contains a set of files organized by L3 domain and by RDT event. E.g. on a system with two L3 domains there will be subdirectories "mon_L3_00" and "mon_L3_01". Each of these directories have one file per event (e.g. "llc_occupancy", "mbm_total_bytes", and "mbm_local_bytes"). In a MON group these files provide a read out of the current value of the event for all tasks in the group. In CTRL_MON groups these files provide the sum for all tasks in the CTRL_MON group and all tasks in MON groups. Please see example section for more details on usage. Resource allocation rules ------------------------- When a task is running the following rules define which resources are available to it: 1) If the task is a member of a non-default group, then the schemata for that group is used. 2) Else if the task belongs to the default group, but is running on a CPU that is assigned to some specific group, then the schemata for the CPU's group is used. 3) Otherwise the schemata for the default group is used. Resource monitoring rules ------------------------- 1) If a task is a member of a MON group, or non-default CTRL_MON group then RDT events for the task will be reported in that group. 2) If a task is a member of the default CTRL_MON group, but is running on a CPU that is assigned to some specific group, then the RDT events for the task will be reported in that group. 3) Otherwise RDT events for the task will be reported in the root level "mon_data" group. Notes on cache occupancy monitoring and control ----------------------------------------------- When moving a task from one group to another you should remember that this only affects *new* cache allocations by the task. E.g. you may have a task in a monitor group showing 3 MB of cache occupancy. If you move to a new group and immediately check the occupancy of the old and new groups you will likely see that the old group is still showing 3 MB and the new group zero. When the task accesses locations still in cache from before the move, the h/w does not update any counters. On a busy system you will likely see the occupancy in the old group go down as cache lines are evicted and re-used while the occupancy in the new group rises as the task accesses memory and loads into the cache are counted based on membership in the new group. The same applies to cache allocation control. Moving a task to a group with a smaller cache partition will not evict any cache lines. The process may continue to use them from the old partition. Hardware uses CLOSid(Class of service ID) and an RMID(Resource monitoring ID) to identify a control group and a monitoring group respectively. Each of the resource groups are mapped to these IDs based on the kind of group. The number of CLOSid and RMID are limited by the hardware and hence the creation of a "CTRL_MON" directory may fail if we run out of either CLOSID or RMID and creation of "MON" group may fail if we run out of RMIDs. max_threshold_occupancy - generic concepts ------------------------------------------ Note that an RMID once freed may not be immediately available for use as the RMID is still tagged the cache lines of the previous user of RMID. Hence such RMIDs are placed on limbo list and checked back if the cache occupancy has gone down. If there is a time when system has a lot of limbo RMIDs but which are not ready to be used, user may see an -EBUSY during mkdir. max_threshold_occupancy is a user configurable value to determine the occupancy at which an RMID can be freed. Schemata files - general concepts --------------------------------- Each line in the file describes one resource. The line starts with the name of the resource, followed by specific values to be applied in each of the instances of that resource on the system. Cache IDs --------- On current generation systems there is one L3 cache per socket and L2 caches are generally just shared by the hyperthreads on a core, but this isn't an architectural requirement. We could have multiple separate L3 caches on a socket, multiple cores could share an L2 cache. So instead of using "socket" or "core" to define the set of logical cpus sharing a resource we use a "Cache ID". At a given cache level this will be a unique number across the whole system (but it isn't guaranteed to be a contiguous sequence, there may be gaps). To find the ID for each logical CPU look in /sys/devices/system/cpu/cpu*/cache/index*/id Cache Bit Masks (CBM) --------------------- For cache resources we describe the portion of the cache that is available for allocation using a bitmask. The maximum value of the mask is defined by each cpu model (and may be different for different cache levels). It is found using CPUID, but is also provided in the "info" directory of the resctrl file system in "info/{resource}/cbm_mask". X86 hardware requires that these masks have all the '1' bits in a contiguous block. So 0x3, 0x6 and 0xC are legal 4-bit masks with two bits set, but 0x5, 0x9 and 0xA are not. On a system with a 20-bit mask each bit represents 5% of the capacity of the cache. You could partition the cache into four equal parts with masks: 0x1f, 0x3e0, 0x7c00, 0xf8000. Memory bandwidth(b/w) percentage -------------------------------- For Memory b/w resource, user controls the resource by indicating the percentage of total memory b/w. The minimum bandwidth percentage value for each cpu model is predefined and can be looked up through "info/MB/min_bandwidth". The bandwidth granularity that is allocated is also dependent on the cpu model and can be looked up at "info/MB/bandwidth_gran". The available bandwidth control steps are: min_bw + N * bw_gran. Intermediate values are rounded to the next control step available on the hardware. The bandwidth throttling is a core specific mechanism on some of Intel SKUs. Using a high bandwidth and a low bandwidth setting on two threads sharing a core will result in both threads being throttled to use the low bandwidth. L3 schemata file details (code and data prioritization disabled) ---------------------------------------------------------------- With CDP disabled the L3 schemata format is: L3:=;=;... L3 schemata file details (CDP enabled via mount option to resctrl) ------------------------------------------------------------------ When CDP is enabled L3 control is split into two separate resources so you can specify independent masks for code and data like this: L3data:=;=;... L3code:=;=;... L2 schemata file details ------------------------ L2 cache does not support code and data prioritization, so the schemata format is always: L2:=;=;... Memory b/w Allocation details ----------------------------- Memory b/w domain is L3 cache. MB:=bandwidth0;=bandwidth1;... Reading/writing the schemata file --------------------------------- Reading the schemata file will show the state of all resources on all domains. When writing you only need to specify those values which you wish to change. E.g. # cat schemata L3DATA:0=fffff;1=fffff;2=fffff;3=fffff L3CODE:0=fffff;1=fffff;2=fffff;3=fffff # echo "L3DATA:2=3c0;" > schemata # cat schemata L3DATA:0=fffff;1=fffff;2=3c0;3=fffff L3CODE:0=fffff;1=fffff;2=fffff;3=fffff Examples for RDT allocation usage: Example 1 --------- On a two socket machine (one L3 cache per socket) with just four bits for cache bit masks, minimum b/w of 10% with a memory bandwidth granularity of 10% # mount -t resctrl resctrl /sys/fs/resctrl # cd /sys/fs/resctrl # mkdir p0 p1 # echo "L3:0=3;1=c\nMB:0=50;1=50" > /sys/fs/resctrl/p0/schemata # echo "L3:0=3;1=3\nMB:0=50;1=50" > /sys/fs/resctrl/p1/schemata The default resource group is unmodified, so we have access to all parts of all caches (its schemata file reads "L3:0=f;1=f"). Tasks that are under the control of group "p0" may only allocate from the "lower" 50% on cache ID 0, and the "upper" 50% of cache ID 1. Tasks in group "p1" use the "lower" 50% of cache on both sockets. Similarly, tasks that are under the control of group "p0" may use a maximum memory b/w of 50% on socket0 and 50% on socket 1. Tasks in group "p1" may also use 50% memory b/w on both sockets. Note that unlike cache masks, memory b/w cannot specify whether these allocations can overlap or not. The allocations specifies the maximum b/w that the group may be able to use and the system admin can configure the b/w accordingly. Example 2 --------- Again two sockets, but this time with a more realistic 20-bit mask. Two real time tasks pid=1234 running on processor 0 and pid=5678 running on processor 1 on socket 0 on a 2-socket and dual core machine. To avoid noisy neighbors, each of the two real-time tasks exclusively occupies one quarter of L3 cache on socket 0. # mount -t resctrl resctrl /sys/fs/resctrl # cd /sys/fs/resctrl First we reset the schemata for the default group so that the "upper" 50% of the L3 cache on socket 0 and 50% of memory b/w cannot be used by ordinary tasks: # echo "L3:0=3ff;1=fffff\nMB:0=50;1=100" > schemata Next we make a resource group for our first real time task and give it access to the "top" 25% of the cache on socket 0. # mkdir p0 # echo "L3:0=f8000;1=fffff" > p0/schemata Finally we move our first real time task into this resource group. We also use taskset(1) to ensure the task always runs on a dedicated CPU on socket 0. Most uses of resource groups will also constrain which processors tasks run on. # echo 1234 > p0/tasks # taskset -cp 1 1234 Ditto for the second real time task (with the remaining 25% of cache): # mkdir p1 # echo "L3:0=7c00;1=fffff" > p1/schemata # echo 5678 > p1/tasks # taskset -cp 2 5678 For the same 2 socket system with memory b/w resource and CAT L3 the schemata would look like(Assume min_bandwidth 10 and bandwidth_gran is 10): For our first real time task this would request 20% memory b/w on socket 0. # echo -e "L3:0=f8000;1=fffff\nMB:0=20;1=100" > p0/schemata For our second real time task this would request an other 20% memory b/w on socket 0. # echo -e "L3:0=f8000;1=fffff\nMB:0=20;1=100" > p0/schemata Example 3 --------- A single socket system which has real-time tasks running on core 4-7 and non real-time workload assigned to core 0-3. The real-time tasks share text and data, so a per task association is not required and due to interaction with the kernel it's desired that the kernel on these cores shares L3 with the tasks. # mount -t resctrl resctrl /sys/fs/resctrl # cd /sys/fs/resctrl First we reset the schemata for the default group so that the "upper" 50% of the L3 cache on socket 0, and 50% of memory bandwidth on socket 0 cannot be used by ordinary tasks: # echo "L3:0=3ff\nMB:0=50" > schemata Next we make a resource group for our real time cores and give it access to the "top" 50% of the cache on socket 0 and 50% of memory bandwidth on socket 0. # mkdir p0 # echo "L3:0=ffc00\nMB:0=50" > p0/schemata Finally we move core 4-7 over to the new group and make sure that the kernel and the tasks running there get 50% of the cache. They should also get 50% of memory bandwidth assuming that the cores 4-7 are SMT siblings and only the real time threads are scheduled on the cores 4-7. # echo F0 > p0/cpus 4) Locking between applications Certain operations on the resctrl filesystem, composed of read/writes to/from multiple files, must be atomic. As an example, the allocation of an exclusive reservation of L3 cache involves: 1. Read the cbmmasks from each directory 2. Find a contiguous set of bits in the global CBM bitmask that is clear in any of the directory cbmmasks 3. Create a new directory 4. Set the bits found in step 2 to the new directory "schemata" file If two applications attempt to allocate space concurrently then they can end up allocating the same bits so the reservations are shared instead of exclusive. To coordinate atomic operations on the resctrlfs and to avoid the problem above, the following locking procedure is recommended: Locking is based on flock, which is available in libc and also as a shell script command Write lock: A) Take flock(LOCK_EX) on /sys/fs/resctrl B) Read/write the directory structure. C) funlock Read lock: A) Take flock(LOCK_SH) on /sys/fs/resctrl B) If success read the directory structure. C) funlock Example with bash: # Atomically read directory structure $ flock -s /sys/fs/resctrl/ find /sys/fs/resctrl # Read directory contents and create new subdirectory $ cat create-dir.sh find /sys/fs/resctrl/ > output.txt mask = function-of(output.txt) mkdir /sys/fs/resctrl/newres/ echo mask > /sys/fs/resctrl/newres/schemata $ flock /sys/fs/resctrl/ ./create-dir.sh Example with C: /* * Example code do take advisory locks * before accessing resctrl filesystem */ #include #include void resctrl_take_shared_lock(int fd) { int ret; /* take shared lock on resctrl filesystem */ ret = flock(fd, LOCK_SH); if (ret) { perror("flock"); exit(-1); } } void resctrl_take_exclusive_lock(int fd) { int ret; /* release lock on resctrl filesystem */ ret = flock(fd, LOCK_EX); if (ret) { perror("flock"); exit(-1); } } void resctrl_release_lock(int fd) { int ret; /* take shared lock on resctrl filesystem */ ret = flock(fd, LOCK_UN); if (ret) { perror("flock"); exit(-1); } } void main(void) { int fd, ret; fd = open("/sys/fs/resctrl", O_DIRECTORY); if (fd == -1) { perror("open"); exit(-1); } resctrl_take_shared_lock(fd); /* code to read directory contents */ resctrl_release_lock(fd); resctrl_take_exclusive_lock(fd); /* code to read and write directory contents */ resctrl_release_lock(fd); } Examples for RDT Monitoring along with allocation usage: Reading monitored data ---------------------- Reading an event file (for ex: mon_data/mon_L3_00/llc_occupancy) would show the current snapshot of LLC occupancy of the corresponding MON group or CTRL_MON group. Example 1 (Monitor CTRL_MON group and subset of tasks in CTRL_MON group) --------- On a two socket machine (one L3 cache per socket) with just four bits for cache bit masks # mount -t resctrl resctrl /sys/fs/resctrl # cd /sys/fs/resctrl # mkdir p0 p1 # echo "L3:0=3;1=c" > /sys/fs/resctrl/p0/schemata # echo "L3:0=3;1=3" > /sys/fs/resctrl/p1/schemata # echo 5678 > p1/tasks # echo 5679 > p1/tasks The default resource group is unmodified, so we have access to all parts of all caches (its schemata file reads "L3:0=f;1=f"). Tasks that are under the control of group "p0" may only allocate from the "lower" 50% on cache ID 0, and the "upper" 50% of cache ID 1. Tasks in group "p1" use the "lower" 50% of cache on both sockets. Create monitor groups and assign a subset of tasks to each monitor group. # cd /sys/fs/resctrl/p1/mon_groups # mkdir m11 m12 # echo 5678 > m11/tasks # echo 5679 > m12/tasks fetch data (data shown in bytes) # cat m11/mon_data/mon_L3_00/llc_occupancy 16234000 # cat m11/mon_data/mon_L3_01/llc_occupancy 14789000 # cat m12/mon_data/mon_L3_00/llc_occupancy 16789000 The parent ctrl_mon group shows the aggregated data. # cat /sys/fs/resctrl/p1/mon_data/mon_l3_00/llc_occupancy 31234000 Example 2 (Monitor a task from its creation) --------- On a two socket machine (one L3 cache per socket) # mount -t resctrl resctrl /sys/fs/resctrl # cd /sys/fs/resctrl # mkdir p0 p1 An RMID is allocated to the group once its created and hence the below is monitored from its creation. # echo $$ > /sys/fs/resctrl/p1/tasks # Fetch the data # cat /sys/fs/resctrl/p1/mon_data/mon_l3_00/llc_occupancy 31789000 Example 3 (Monitor without CAT support or before creating CAT groups) --------- Assume a system like HSW has only CQM and no CAT support. In this case the resctrl will still mount but cannot create CTRL_MON directories. But user can create different MON groups within the root group thereby able to monitor all tasks including kernel threads. This can also be used to profile jobs cache size footprint before being able to allocate them to different allocation groups. # mount -t resctrl resctrl /sys/fs/resctrl # cd /sys/fs/resctrl # mkdir mon_groups/m01 # mkdir mon_groups/m02 # echo 3478 > /sys/fs/resctrl/mon_groups/m01/tasks # echo 2467 > /sys/fs/resctrl/mon_groups/m02/tasks Monitor the groups separately and also get per domain data. From the below its apparent that the tasks are mostly doing work on domain(socket) 0. # cat /sys/fs/resctrl/mon_groups/m01/mon_L3_00/llc_occupancy 31234000 # cat /sys/fs/resctrl/mon_groups/m01/mon_L3_01/llc_occupancy 34555 # cat /sys/fs/resctrl/mon_groups/m02/mon_L3_00/llc_occupancy 31234000 # cat /sys/fs/resctrl/mon_groups/m02/mon_L3_01/llc_occupancy 32789 Example 4 (Monitor real time tasks) ----------------------------------- A single socket system which has real time tasks running on cores 4-7 and non real time tasks on other cpus. We want to monitor the cache occupancy of the real time threads on these cores. # mount -t resctrl resctrl /sys/fs/resctrl # cd /sys/fs/resctrl # mkdir p1 Move the cpus 4-7 over to p1 # echo f0 > p0/cpus View the llc occupancy snapshot # cat /sys/fs/resctrl/p1/mon_data/mon_L3_00/llc_occupancy 11234000