Re: [Qemu-devel] [PATCH V11 1/5] throttle: Add a new throttling API implementing continuous leaky bucket.

[Fam Zheng]

On Sun, 09/01 18:39, Benoît Canet wrote:
> Implement the continuous leaky bucket algorithm devised on IRC as a separate
> module.
> 
> Signed-off-by: Benoit Canet <address@hidden>
> ---
>  include/qemu/throttle.h |  103 ++++++++++++
>  util/Makefile.objs      |    1 +
>  util/throttle.c         |  396 
> +++++++++++++++++++++++++++++++++++++++++++++++
>  3 files changed, 500 insertions(+)
>  create mode 100644 include/qemu/throttle.h
>  create mode 100644 util/throttle.c
> 
> diff --git a/include/qemu/throttle.h b/include/qemu/throttle.h
> new file mode 100644
> index 0000000..823650d
> --- /dev/null
> +++ b/include/qemu/throttle.h
> @@ -0,0 +1,103 @@
> +/*
> + * QEMU throttling infrastructure
> + *
> + * Copyright (C) Nodalink, SARL. 2013
> + *
> + * Author:
> + *   Benoît Canet <address@hidden>
> + *
> + * This program is free software; you can redistribute it and/or
> + * modify it under the terms of the GNU General Public License as
> + * published by the Free Software Foundation; either version 2 or
> + * (at your option) version 3 of the License.
> + *
> + * This program is distributed in the hope that it will be useful,
> + * but WITHOUT ANY WARRANTY; without even the implied warranty of
> + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
> + * GNU General Public License for more details.
> + *
> + * You should have received a copy of the GNU General Public License
> + * along with this program; if not, see <http://www.gnu.org/licenses/>.
> + */
> +
> +#ifndef THROTTLE_H
> +#define THROTTLE_H
> +
> +#include <stdint.h>
> +#include "qemu-common.h"
> +#include "qemu/timer.h"
> +
> +#define NANOSECONDS_PER_SECOND  1000000000.0
> +
> +typedef enum {
> +    THROTTLE_BPS_TOTAL,
> +    THROTTLE_BPS_READ,
> +    THROTTLE_BPS_WRITE,
> +    THROTTLE_OPS_TOTAL,
> +    THROTTLE_OPS_READ,
> +    THROTTLE_OPS_WRITE,
> +    BUCKETS_COUNT,
> +} BucketType;
> +
> +typedef struct LeakyBucket {
> +    double  avg;              /* average goal in units per second */
> +    double  max;              /* leaky bucket max burst in units */
> +    double  level;            /* bucket level in units */
> +} LeakyBucket;
> +
> +/* The following structure is used to configure a ThrottleState
> + * It contains a bit of state: the bucket field of the LeakyBucket structure.
> + * However it allows to keep the code clean and the bucket field is reset to
> + * zero at the right time.
> + */
> +typedef struct ThrottleConfig {
> +    LeakyBucket buckets[BUCKETS_COUNT]; /* leaky buckets */
> +    uint64_t op_size;         /* size of an operation in bytes */
> +} ThrottleConfig;
> +
> +typedef struct ThrottleState {
> +    ThrottleConfig cfg;       /* configuration */
> +    int64_t previous_leak;    /* timestamp of the last leak done */
> +    QEMUTimer * timers[2];    /* timers used to do the throttling */
> +    QEMUClockType clock_type; /* the clock used */
> +} ThrottleState;
> +
> +/* operations on single leaky buckets */
> +void throttle_leak_bucket(LeakyBucket *bkt, int64_t delta);
> +
> +int64_t throttle_compute_wait(LeakyBucket *bkt);
> +
> +/* expose timer computation function for unit tests */
> +bool throttle_compute_timer(ThrottleState *ts,
> +                            bool is_write,
> +                            int64_t now,
> +                            int64_t *next_timestamp);
> +
> +/* init/destroy cycle */
> +void throttle_init(ThrottleState *ts,
> +                   QEMUClockType clock_type,
> +                   void (read_timer)(void *),
> +                   void (write_timer)(void *),
> +                   void *timer_opaque);
> +
> +void throttle_destroy(ThrottleState *ts);
> +
> +bool throttle_have_timer(ThrottleState *ts);
> +
> +/* configuration */
> +bool throttle_enabled(ThrottleConfig *cfg);
> +
> +bool throttle_conflicting(ThrottleConfig *cfg);
> +
> +bool throttle_is_valid(ThrottleConfig *cfg);
> +
> +void throttle_config(ThrottleState *ts, ThrottleConfig *cfg);
> +
> +void throttle_get_config(ThrottleState *ts, ThrottleConfig *cfg);
> +
> +/* usage */
> +bool throttle_schedule_timer(ThrottleState *ts, bool is_write);
> +
> +void throttle_account(ThrottleState *ts, bool is_write, uint64_t size);
> +
> +#endif
> diff --git a/util/Makefile.objs b/util/Makefile.objs
> index dc72ab0..2bb13a2 100644
> --- a/util/Makefile.objs
> +++ b/util/Makefile.objs
> @@ -11,3 +11,4 @@ util-obj-y += iov.o aes.o qemu-config.o qemu-sockets.o 
> uri.o notify.o
>  util-obj-y += qemu-option.o qemu-progress.o
>  util-obj-y += hexdump.o
>  util-obj-y += crc32c.o
> +util-obj-y += throttle.o
> diff --git a/util/throttle.c b/util/throttle.c
> new file mode 100644
> index 0000000..cf048b9
> --- /dev/null
> +++ b/util/throttle.c
> @@ -0,0 +1,396 @@
> +/*
> + * QEMU throttling infrastructure
> + *
> + * Copyright (C) Nodalink, SARL. 2013
> + *
> + * Author:
> + *   Benoît Canet <address@hidden>
> + *
> + * This program is free software; you can redistribute it and/or
> + * modify it under the terms of the GNU General Public License as
> + * published by the Free Software Foundation; either version 2 or
> + * (at your option) version 3 of the License.
> + *
> + * This program is distributed in the hope that it will be useful,
> + * but WITHOUT ANY WARRANTY; without even the implied warranty of
> + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
> + * GNU General Public License for more details.
> + *
> + * You should have received a copy of the GNU General Public License
> + * along with this program; if not, see <http://www.gnu.org/licenses/>.
> + */
> +
> +#include "qemu/throttle.h"
> +#include "qemu/timer.h"
> +
> +/* This function make a bucket leak
> + *
> + * @bkt:   the bucket to make leak
> + * @delta_ns: the time delta
> + */
> +void throttle_leak_bucket(LeakyBucket *bkt, int64_t delta_ns)
> +{
> +    double leak;
> +
> +    /* compute how much to leak */
> +    leak = (bkt->avg * (double) delta_ns) / NANOSECONDS_PER_SECOND;
> +
> +    /* make the bucket leak */
> +    bkt->level = MAX(bkt->level - leak, 0);
> +}
> +
> +/* Calculate the time delta since last leak and make proportionals leaks
> + *
> + * @now:      the current timestamp in ns
> + */
> +static void throttle_do_leak(ThrottleState *ts, int64_t now)
> +{
> +    /* compute the time elapsed since the last leak */
> +    int64_t delta_ns = now - ts->previous_leak;
> +    int i;
> +
> +    ts->previous_leak = now;
> +
> +    if (delta_ns <= 0) {
> +        return;
> +    }
> +
> +    /* make each bucket leak */
> +    for (i = 0; i < BUCKETS_COUNT; i++) {
> +        throttle_leak_bucket(&ts->cfg.buckets[i], delta_ns);
> +    }
> +}
> +
> +/* do the real job of computing the time to wait
> + *
> + * @limit: the throttling limit
> + * @extra: the number of operation to delay
> + * @ret:   the time to wait in ns
> + */
> +static int64_t throttle_do_compute_wait(double limit, double extra)
> +{
> +    double wait = extra * NANOSECONDS_PER_SECOND;
> +    wait /= limit;
> +    return wait;
> +}
> +
> +/* This function compute the wait time in ns that a leaky bucket should 
> trigger
> + *
> + * @bkt: the leaky bucket we operate on
> + * @ret: the resulting wait time in ns or 0 if the operation can go through
> + */
> +int64_t throttle_compute_wait(LeakyBucket *bkt)
> +{
> +    double extra; /* the number of extra units blocking the io */
> +
> +    if (!bkt->avg) {
> +        return 0;
> +    }
> +
> +    extra = bkt->level - bkt->max;
> +
> +    if (extra <= 0) {
> +        return 0;
> +    }
> +
> +    return throttle_do_compute_wait(bkt->avg, extra);
> +}
> +
> +/* This function compute the time that must be waited while this IO
> + *
> + * @is_write:   true if the current IO is a write, false if it's a read
> + * @ret:        time to wait
> + */
> +static int64_t throttle_compute_wait_for(ThrottleState *ts,
> +                                         bool is_write)
> +{
> +    BucketType to_check[2][4] = { {THROTTLE_BPS_TOTAL,
> +                                   THROTTLE_OPS_TOTAL,
> +                                   THROTTLE_BPS_READ,
> +                                   THROTTLE_OPS_READ},
> +                                  {THROTTLE_BPS_TOTAL,
> +                                   THROTTLE_OPS_TOTAL,
> +                                   THROTTLE_BPS_WRITE,
> +                                   THROTTLE_OPS_WRITE}, };
> +    int64_t wait, max_wait = 0;
> +    int i;
> +
> +    for (i = 0; i < 4; i++) {
> +        BucketType index = to_check[is_write][i];
> +        wait = throttle_compute_wait(&ts->cfg.buckets[index]);
> +        if (wait > max_wait) {
> +            max_wait = wait;
> +        }
> +    }
> +
> +    return max_wait;
> +}
> +
> +/* compute the timer for this type of operation
> + *
> + * @is_write:   the type of operation
> + * @now:        the current clock timestamp
> + * @next_timestamp: the resulting timer
> + * @ret:        true if a timer must be set
> + */
> +bool throttle_compute_timer(ThrottleState *ts,
> +                            bool is_write,
> +                            int64_t now,
> +                            int64_t *next_timestamp)
> +{
> +    int64_t wait;
> +
> +    /* leak proportionally to the time elapsed */
> +    throttle_do_leak(ts, now);
> +
> +    /* compute the wait time if any */
> +    wait = throttle_compute_wait_for(ts, is_write);
> +
> +    /* if the code must wait compute when the next timer should fire */
> +    if (wait) {
> +        *next_timestamp = now + wait;
> +        return true;
> +    }
> +
> +    /* else no need to wait at all */
> +    *next_timestamp = now;
> +    return false;
> +}
> +
> +/* To be called first on the ThrottleState */
> +void throttle_init(ThrottleState *ts,
> +                   QEMUClockType clock_type,
> +                   QEMUTimerCB *read_timer_cb,
> +                   QEMUTimerCB *write_timer_cb,
> +                   void *timer_opaque)
> +{
> +    memset(ts, 0, sizeof(ThrottleState));
> +
> +    ts->clock_type = clock_type;
> +    ts->timers[0] = timer_new_ns(clock_type, read_timer_cb, timer_opaque);
> +    ts->timers[1] = timer_new_ns(clock_type, write_timer_cb, timer_opaque);
> +}
> +
> +/* destroy a timer */
> +static void throttle_timer_destroy(QEMUTimer **timer)
> +{
> +    assert(*timer != NULL);
> +
> +    timer_del(*timer);
> +    timer_free(*timer);
> +    *timer = NULL;
> +}
> +
> +/* To be called last on the ThrottleState */
> +void throttle_destroy(ThrottleState *ts)
> +{
> +    int i;
> +
> +    for (i = 0; i < 2; i++) {
> +        throttle_timer_destroy(&ts->timers[i]);
> +    }
> +}
> +
> +/* is any throttling timer configured */
> +bool throttle_have_timer(ThrottleState *ts)
> +{
> +    if (ts->timers[0]) {
> +        return true;
> +    }
> +
> +    return false;
> +}
> +
> +/* Does any throttling must be done
> + *
> + * @cfg: the throttling configuration to inspect
> + * @ret: true if throttling must be done else false
> + */
> +bool throttle_enabled(ThrottleConfig *cfg)
> +{
> +    int i;
> +
> +    for (i = 0; i < BUCKETS_COUNT; i++) {
> +        if (cfg->buckets[i].avg > 0) {
> +            return true;
> +        }
> +    }
> +
> +    return false;
> +}
> +
> +/* return true if any two throttling parameters conflicts
> + *
> + * @cfg: the throttling configuration to inspect
> + * @ret: true if any conflict detected else false
> + */
> +bool throttle_conflicting(ThrottleConfig *cfg)
> +{
> +    bool bps_flag, ops_flag;
> +    bool bps_max_flag, ops_max_flag;
> +
> +    bps_flag = cfg->buckets[THROTTLE_BPS_TOTAL].avg &&
> +               (cfg->buckets[THROTTLE_BPS_READ].avg ||
> +                cfg->buckets[THROTTLE_BPS_WRITE].avg);
> +
> +    ops_flag = cfg->buckets[THROTTLE_OPS_TOTAL].avg &&
> +               (cfg->buckets[THROTTLE_OPS_READ].avg ||
> +                cfg->buckets[THROTTLE_OPS_WRITE].avg);
> +
> +    bps_max_flag = cfg->buckets[THROTTLE_BPS_TOTAL].max &&
> +                  (cfg->buckets[THROTTLE_BPS_READ].max  ||
> +                   cfg->buckets[THROTTLE_BPS_WRITE].max);
> +
> +    ops_max_flag = cfg->buckets[THROTTLE_OPS_TOTAL].max &&
> +                   (cfg->buckets[THROTTLE_OPS_READ].max ||
> +                   cfg->buckets[THROTTLE_OPS_WRITE].max);
> +
> +    return bps_flag || ops_flag || bps_max_flag || ops_max_flag;
> +}
> +
> +/* check if a throttling configuration is valid
> + * @cfg: the throttling configuration to inspect
> + * @ret: true if valid else false
> + */
> +bool throttle_is_valid(ThrottleConfig *cfg)
> +{
> +    bool invalid = false;
> +    int i;
> +
> +    for (i = 0; i < BUCKETS_COUNT; i++) {
> +        if (cfg->buckets[i].avg < 0) {
> +            invalid = true;
> +        }
> +    }
> +
> +    for (i = 0; i < BUCKETS_COUNT; i++) {
> +        if (cfg->buckets[i].max < 0) {
> +            invalid = true;
> +        }
> +    }
> +
> +    return !invalid;
> +}
> +
> +/* fix bucket parameters */
> +static void throttle_fix_bucket(LeakyBucket *bkt)
> +{
> +    double min;
> +
> +    /* zero bucket level */
> +    bkt->level = 0;
> +
> +    /* The following is done to cope with the Linux CFQ block scheduler
> +     * which regroup reads and writes by block of 100ms in the guest.
> +     * When they are two process one making reads and one making writes cfq
> +     * make a pattern looking like the following:
> +     * WWWWWWWWWWWRRRRRRRRRRRRRRWWWWWWWWWWWWWwRRRRRRRRRRRRRRRRR
> +     * Having a max burst value of 100ms of the average will help smooth the
> +     * throttling
> +     */
> +    min = bkt->avg / 10;
> +    if (bkt->avg && !bkt->max) {
> +        bkt->max = min;
> +    }
> +}
> +
> +/* take care of canceling a timer */
> +static void throttle_cancel_timer(QEMUTimer *timer)
> +{
> +    assert(timer != NULL);
> +
> +    timer_del(timer);
> +}
> +
> +/* Used to configure the throttle
> + *
> + * @ts: the throttle state we are working on
> + * @cfg: the config to set
> + */
> +void throttle_config(ThrottleState *ts, ThrottleConfig *cfg)
> +{
> +    int i;
> +
> +    ts->cfg = *cfg;
> +
> +    for (i = 0; i < BUCKETS_COUNT; i++) {
> +        throttle_fix_bucket(&ts->cfg.buckets[i]);
> +    }
> +
> +    ts->previous_leak = qemu_clock_get_ns(ts->clock_type);
> +
> +    for (i = 0; i < 2; i++) {
> +        throttle_cancel_timer(ts->timers[i]);
> +    }
> +}
> +
> +/* used to get config
> + *
> + * @ts:  the throttle state we are working on
> + * @cfg: the config to write
> + */
> +void throttle_get_config(ThrottleState *ts, ThrottleConfig *cfg)
> +{
> +    *cfg = ts->cfg;
> +}
> +
> +
> +/* Schedule the read or write timer if needed
> + *
> + * NOTE: this function is not unit tested due to it's usage of timer_mod
> + *
> + * @is_write: the type of operation (read/write)
> + * @ret:      true if the timer has been scheduled else false
> + */
> +bool throttle_schedule_timer(ThrottleState *ts, bool is_write)
> +{
> +    int64_t now = qemu_clock_get_ns(ts->clock_type);
> +    int64_t next_timestamp;
> +    bool must_wait;
> +
> +    must_wait = throttle_compute_timer(ts,
> +                                       is_write,
> +                                       now,
> +                                       &next_timestamp);
> +
> +    /* request not throttled */
> +    if (!must_wait) {
> +        return false;
> +    }
> +
> +    /* request throttled and timer pending -> do nothing */
> +    if (timer_pending(ts->timers[is_write])) {
> +        return true;
> +    }
> +
> +    /* request throttled and timer not pending -> arm timer */
> +    timer_mod(ts->timers[is_write], next_timestamp);
> +    return true;
> +}
> +
> +/* do the accounting for this operation
> + *
> + * @is_write: the type of operation (read/write)
> + * @size:     the size of the operation
> + */
> +void throttle_account(ThrottleState *ts, bool is_write, uint64_t size)
> +{
> +    double units = 1.0;
> +
> +    /* if cfg.op_size is not defined we will account exactly 1 operation */
> +    if (ts->cfg.op_size) {
> +        units = (double) size / ts->cfg.op_size;
> +    }

If op_size is non-zero, iops limits are merely a fixed proportion of bps
limits, which means the lower set of the two is applied and the higher skipped.
I understand the amazon uses op_size like accounting for big IO requests, but
we don't do it condionally on io size or anything here, so that once user sets
op_size, it simply kicks either bps_{,rd,wr} or iops_{,rd,wr} out the game, is
that true?

Fam
> +
> +    ts->cfg.buckets[THROTTLE_BPS_TOTAL].level += size;
> +    ts->cfg.buckets[THROTTLE_OPS_TOTAL].level += units;
> +
> +    if (is_write) {
> +        ts->cfg.buckets[THROTTLE_BPS_WRITE].level += size;
> +        ts->cfg.buckets[THROTTLE_OPS_WRITE].level += units;
> +    } else {
> +        ts->cfg.buckets[THROTTLE_BPS_READ].level += size;
> +        ts->cfg.buckets[THROTTLE_OPS_READ].level += units;
> +    }
> +}
> +
> -- 
> 1.7.10.4
> 
>