cpufreq: powernv: Ramp-down global pstate slower than local-pstate
The frequency transition latency from pmin to pmax is observed to be in
few millisecond granurality. And it usually happens to take a performance
penalty during sudden frequency rampup requests.
This patch set solves this problem by using an entity called "global
pstates". The global pstate is a Chip-level entity, so the global entitiy
(Voltage) is managed across the cores. The local pstate is a Core-level
entity, so the local entity (frequency) is managed across threads.
This patch brings down global pstate at a slower rate than the local
pstate. Hence by holding global pstates higher than local pstate makes
the subsequent rampups faster.
A per policy structure is maintained to keep track of the global and
local pstate changes. The global pstate is brought down using a parabolic
equation. The ramp down time to pmin is set to ~5 seconds. To make sure
that the global pstates are dropped at regular interval , a timer is
queued for every 2 seconds during ramp-down phase, which eventually brings
the pstate down to local pstate.
Iozone results show fairly consistent performance boost.
YCSB on redis shows improved Max latencies in most cases.
Iozone write/rewite test were made with filesizes 200704Kb and 401408Kb
with different record sizes . The following table shows IOoperations/sec
with and without patch.
Iozone Results ( in op/sec) ( mean over 3 iterations )
---------------------------------------------------------------------
file size- with without %
recordsize-IOtype patch patch change
----------------------------------------------------------------------
200704-1-SeqWrite 1616532 1615425 0.06
200704-1-Rewrite 2423195 2303130 5.21
200704-2-SeqWrite 1628577 1602620 1.61
200704-2-Rewrite 2428264 2312154 5.02
200704-4-SeqWrite 1617605 1617182 0.02
200704-4-Rewrite 2430524 2351238 3.37
200704-8-SeqWrite 1629478 1600436 1.81
200704-8-Rewrite 2415308 2298136 5.09
200704-16-SeqWrite 1619632 1618250 0.08
200704-16-Rewrite 2396650 2352591 1.87
200704-32-SeqWrite 1632544 1598083 2.15
200704-32-Rewrite 2425119 2329743 4.09
200704-64-SeqWrite 1617812 1617235 0.03
200704-64-Rewrite 2402021 2321080 3.48
200704-128-SeqWrite 1631998 1600256 1.98
200704-128-Rewrite 2422389 2304954 5.09
200704-256 SeqWrite 1617065 1616962 0.00
200704-256-Rewrite 2432539 2301980 5.67
200704-512-SeqWrite 1632599 1598656 2.12
200704-512-Rewrite 2429270 2323676 4.54
200704-1024-SeqWrite 1618758 1616156 0.16
200704-1024-Rewrite 2431631 2315889 4.99
401408-1-SeqWrite 1631479 1608132 1.45
401408-1-Rewrite 2501550 2459409 1.71
401408-2-SeqWrite 1617095 1626069 -0.55
401408-2-Rewrite 2507557 2443621 2.61
401408-4-SeqWrite 1629601 1611869 1.10
401408-4-Rewrite 2505909 2462098 1.77
401408-8-SeqWrite 1617110 1626968 -0.60
401408-8-Rewrite 2512244 2456827 2.25
401408-16-SeqWrite 1632609 1609603 1.42
401408-16-Rewrite 2500792 2451405 2.01
401408-32-SeqWrite 1619294 1628167 -0.54
401408-32-Rewrite 2510115 2451292 2.39
401408-64-SeqWrite 1632709 1603746 1.80
401408-64-Rewrite 2506692 2433186 3.02
401408-128-SeqWrite 1619284 1627461 -0.50
401408-128-Rewrite 2518698 2453361 2.66
401408-256-SeqWrite 1634022 1610681 1.44
401408-256-Rewrite 2509987 2446328 2.60
401408-512-SeqWrite 1617524 1628016 -0.64
401408-512-Rewrite 2504409 2442899 2.51
401408-1024-SeqWrite 1629812 1611566 1.13
401408-1024-Rewrite 2507620 2442968 2.64
Tested with YCSB workload (50% update + 50% read) over redis for 1 million
records and 1 million operation. Each test was carried out with target
operations per second and persistence disabled.
Max-latency (in us)( mean over 5 iterations )
---------------------------------------------------------------
op/s Operation with patch without patch %change
---------------------------------------------------------------
15000 Read 61480.6 50261.4 22.32
15000 cleanup 215.2 293.6 -26.70
15000 update 25666.2 25163.8 2.00
25000 Read 32626.2 89525.4 -63.56
25000 cleanup 292.2 263.0 11.10
25000 update 32293.4 90255.0 -64.22
35000 Read 34783.0 33119.0 5.02
35000 cleanup 321.2 395.8 -18.8
35000 update 36047.0 38747.8 -6.97
40000 Read 38562.2 42357.4 -8.96
40000 cleanup 371.8 384.6 -3.33
40000 update 27861.4 41547.8 -32.94
45000 Read 42271.0 88120.6 -52.03
45000 cleanup 263.6 383.0 -31.17
45000 update 29755.8 81359.0 -63.43
(test without target op/s)
47659 Read 83061.4 136440.6 -39.12
47659 cleanup 195.8 193.8 1.03
47659 update 73429.4 124971.8 -41.24
Signed-off-by: Akshay Adiga <akshay.adiga@linux.vnet.ibm.com>
Reviewed-by: Gautham R. Shenoy <ego@linux.vnet.ibm.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
diff --git a/drivers/cpufreq/powernv-cpufreq.c b/drivers/cpufreq/powernv-cpufreq.c
index e2e2219..144c732 100644
--- a/drivers/cpufreq/powernv-cpufreq.c
+++ b/drivers/cpufreq/powernv-cpufreq.c
@@ -36,12 +36,56 @@
#include <asm/reg.h>
#include <asm/smp.h> /* Required for cpu_sibling_mask() in UP configs */
#include <asm/opal.h>
+#include <linux/timer.h>
#define POWERNV_MAX_PSTATES 256
#define PMSR_PSAFE_ENABLE (1UL << 30)
#define PMSR_SPR_EM_DISABLE (1UL << 31)
#define PMSR_MAX(x) ((x >> 32) & 0xFF)
+#define MAX_RAMP_DOWN_TIME 5120
+/*
+ * On an idle system we want the global pstate to ramp-down from max value to
+ * min over a span of ~5 secs. Also we want it to initially ramp-down slowly and
+ * then ramp-down rapidly later on.
+ *
+ * This gives a percentage rampdown for time elapsed in milliseconds.
+ * ramp_down_percentage = ((ms * ms) >> 18)
+ * ~= 3.8 * (sec * sec)
+ *
+ * At 0 ms ramp_down_percent = 0
+ * At 5120 ms ramp_down_percent = 100
+ */
+#define ramp_down_percent(time) ((time * time) >> 18)
+
+/* Interval after which the timer is queued to bring down global pstate */
+#define GPSTATE_TIMER_INTERVAL 2000
+
+/**
+ * struct global_pstate_info - Per policy data structure to maintain history of
+ * global pstates
+ * @highest_lpstate: The local pstate from which we are ramping down
+ * @elapsed_time: Time in ms spent in ramping down from
+ * highest_lpstate
+ * @last_sampled_time: Time from boot in ms when global pstates were
+ * last set
+ * @last_lpstate,last_gpstate: Last set values for local and global pstates
+ * @timer: Is used for ramping down if cpu goes idle for
+ * a long time with global pstate held high
+ * @gpstate_lock: A spinlock to maintain synchronization between
+ * routines called by the timer handler and
+ * governer's target_index calls
+ */
+struct global_pstate_info {
+ int highest_lpstate;
+ unsigned int elapsed_time;
+ unsigned int last_sampled_time;
+ int last_lpstate;
+ int last_gpstate;
+ spinlock_t gpstate_lock;
+ struct timer_list timer;
+};
+
static struct cpufreq_frequency_table powernv_freqs[POWERNV_MAX_PSTATES+1];
static bool rebooting, throttled, occ_reset;
@@ -94,6 +138,17 @@
int nr_pstates;
} powernv_pstate_info;
+static inline void reset_gpstates(struct cpufreq_policy *policy)
+{
+ struct global_pstate_info *gpstates = policy->driver_data;
+
+ gpstates->highest_lpstate = 0;
+ gpstates->elapsed_time = 0;
+ gpstates->last_sampled_time = 0;
+ gpstates->last_lpstate = 0;
+ gpstates->last_gpstate = 0;
+}
+
/*
* Initialize the freq table based on data obtained
* from the firmware passed via device-tree
@@ -285,6 +340,7 @@
struct powernv_smp_call_data {
unsigned int freq;
int pstate_id;
+ int gpstate_id;
};
/*
@@ -343,19 +399,21 @@
* (struct powernv_smp_call_data *) and the pstate_id which needs to be set
* on this CPU should be present in freq_data->pstate_id.
*/
-static void set_pstate(void *freq_data)
+static void set_pstate(void *data)
{
unsigned long val;
- unsigned long pstate_ul =
- ((struct powernv_smp_call_data *) freq_data)->pstate_id;
+ struct powernv_smp_call_data *freq_data = data;
+ unsigned long pstate_ul = freq_data->pstate_id;
+ unsigned long gpstate_ul = freq_data->gpstate_id;
val = get_pmspr(SPRN_PMCR);
val = val & 0x0000FFFFFFFFFFFFULL;
pstate_ul = pstate_ul & 0xFF;
+ gpstate_ul = gpstate_ul & 0xFF;
/* Set both global(bits 56..63) and local(bits 48..55) PStates */
- val = val | (pstate_ul << 56) | (pstate_ul << 48);
+ val = val | (gpstate_ul << 56) | (pstate_ul << 48);
pr_debug("Setting cpu %d pmcr to %016lX\n",
raw_smp_processor_id(), val);
@@ -424,6 +482,110 @@
}
}
+/**
+ * calc_global_pstate - Calculate global pstate
+ * @elapsed_time: Elapsed time in milliseconds
+ * @local_pstate: New local pstate
+ * @highest_lpstate: pstate from which its ramping down
+ *
+ * Finds the appropriate global pstate based on the pstate from which its
+ * ramping down and the time elapsed in ramping down. It follows a quadratic
+ * equation which ensures that it reaches ramping down to pmin in 5sec.
+ */
+static inline int calc_global_pstate(unsigned int elapsed_time,
+ int highest_lpstate, int local_pstate)
+{
+ int pstate_diff;
+
+ /*
+ * Using ramp_down_percent we get the percentage of rampdown
+ * that we are expecting to be dropping. Difference between
+ * highest_lpstate and powernv_pstate_info.min will give a absolute
+ * number of how many pstates we will drop eventually by the end of
+ * 5 seconds, then just scale it get the number pstates to be dropped.
+ */
+ pstate_diff = ((int)ramp_down_percent(elapsed_time) *
+ (highest_lpstate - powernv_pstate_info.min)) / 100;
+
+ /* Ensure that global pstate is >= to local pstate */
+ if (highest_lpstate - pstate_diff < local_pstate)
+ return local_pstate;
+ else
+ return highest_lpstate - pstate_diff;
+}
+
+static inline void queue_gpstate_timer(struct global_pstate_info *gpstates)
+{
+ unsigned int timer_interval;
+
+ /*
+ * Setting up timer to fire after GPSTATE_TIMER_INTERVAL ms, But
+ * if it exceeds MAX_RAMP_DOWN_TIME ms for ramp down time.
+ * Set timer such that it fires exactly at MAX_RAMP_DOWN_TIME
+ * seconds of ramp down time.
+ */
+ if ((gpstates->elapsed_time + GPSTATE_TIMER_INTERVAL)
+ > MAX_RAMP_DOWN_TIME)
+ timer_interval = MAX_RAMP_DOWN_TIME - gpstates->elapsed_time;
+ else
+ timer_interval = GPSTATE_TIMER_INTERVAL;
+
+ mod_timer_pinned(&gpstates->timer, jiffies +
+ msecs_to_jiffies(timer_interval));
+}
+
+/**
+ * gpstate_timer_handler
+ *
+ * @data: pointer to cpufreq_policy on which timer was queued
+ *
+ * This handler brings down the global pstate closer to the local pstate
+ * according quadratic equation. Queues a new timer if it is still not equal
+ * to local pstate
+ */
+void gpstate_timer_handler(unsigned long data)
+{
+ struct cpufreq_policy *policy = (struct cpufreq_policy *)data;
+ struct global_pstate_info *gpstates = policy->driver_data;
+ int gpstate_id;
+ unsigned int time_diff = jiffies_to_msecs(jiffies)
+ - gpstates->last_sampled_time;
+ struct powernv_smp_call_data freq_data;
+
+ if (!spin_trylock(&gpstates->gpstate_lock))
+ return;
+
+ gpstates->last_sampled_time += time_diff;
+ gpstates->elapsed_time += time_diff;
+ freq_data.pstate_id = gpstates->last_lpstate;
+
+ if ((gpstates->last_gpstate == freq_data.pstate_id) ||
+ (gpstates->elapsed_time > MAX_RAMP_DOWN_TIME)) {
+ gpstate_id = freq_data.pstate_id;
+ reset_gpstates(policy);
+ gpstates->highest_lpstate = freq_data.pstate_id;
+ } else {
+ gpstate_id = calc_global_pstate(gpstates->elapsed_time,
+ gpstates->highest_lpstate,
+ freq_data.pstate_id);
+ }
+
+ /*
+ * If local pstate is equal to global pstate, rampdown is over
+ * So timer is not required to be queued.
+ */
+ if (gpstate_id != freq_data.pstate_id)
+ queue_gpstate_timer(gpstates);
+
+ freq_data.gpstate_id = gpstate_id;
+ gpstates->last_gpstate = freq_data.gpstate_id;
+ gpstates->last_lpstate = freq_data.pstate_id;
+
+ /* Timer may get migrated to a different cpu on cpu hot unplug */
+ smp_call_function_any(policy->cpus, set_pstate, &freq_data, 1);
+ spin_unlock(&gpstates->gpstate_lock);
+}
+
/*
* powernv_cpufreq_target_index: Sets the frequency corresponding to
* the cpufreq table entry indexed by new_index on the cpus in the
@@ -433,6 +595,9 @@
unsigned int new_index)
{
struct powernv_smp_call_data freq_data;
+ unsigned int cur_msec, gpstate_id;
+ unsigned long flags;
+ struct global_pstate_info *gpstates = policy->driver_data;
if (unlikely(rebooting) && new_index != get_nominal_index())
return 0;
@@ -440,22 +605,70 @@
if (!throttled)
powernv_cpufreq_throttle_check(NULL);
+ cur_msec = jiffies_to_msecs(get_jiffies_64());
+
+ spin_lock_irqsave(&gpstates->gpstate_lock, flags);
freq_data.pstate_id = powernv_freqs[new_index].driver_data;
+ if (!gpstates->last_sampled_time) {
+ gpstate_id = freq_data.pstate_id;
+ gpstates->highest_lpstate = freq_data.pstate_id;
+ goto gpstates_done;
+ }
+
+ if (gpstates->last_gpstate > freq_data.pstate_id) {
+ gpstates->elapsed_time += cur_msec -
+ gpstates->last_sampled_time;
+
+ /*
+ * If its has been ramping down for more than MAX_RAMP_DOWN_TIME
+ * we should be resetting all global pstate related data. Set it
+ * equal to local pstate to start fresh.
+ */
+ if (gpstates->elapsed_time > MAX_RAMP_DOWN_TIME) {
+ reset_gpstates(policy);
+ gpstates->highest_lpstate = freq_data.pstate_id;
+ gpstate_id = freq_data.pstate_id;
+ } else {
+ /* Elaspsed_time is less than 5 seconds, continue to rampdown */
+ gpstate_id = calc_global_pstate(gpstates->elapsed_time,
+ gpstates->highest_lpstate,
+ freq_data.pstate_id);
+ }
+ } else {
+ reset_gpstates(policy);
+ gpstates->highest_lpstate = freq_data.pstate_id;
+ gpstate_id = freq_data.pstate_id;
+ }
+
+ /*
+ * If local pstate is equal to global pstate, rampdown is over
+ * So timer is not required to be queued.
+ */
+ if (gpstate_id != freq_data.pstate_id)
+ queue_gpstate_timer(gpstates);
+
+gpstates_done:
+ freq_data.gpstate_id = gpstate_id;
+ gpstates->last_sampled_time = cur_msec;
+ gpstates->last_gpstate = freq_data.gpstate_id;
+ gpstates->last_lpstate = freq_data.pstate_id;
+
/*
* Use smp_call_function to send IPI and execute the
* mtspr on target CPU. We could do that without IPI
* if current CPU is within policy->cpus (core)
*/
smp_call_function_any(policy->cpus, set_pstate, &freq_data, 1);
-
+ spin_unlock_irqrestore(&gpstates->gpstate_lock, flags);
return 0;
}
static int powernv_cpufreq_cpu_init(struct cpufreq_policy *policy)
{
- int base, i;
+ int base, i, ret;
struct kernfs_node *kn;
+ struct global_pstate_info *gpstates;
base = cpu_first_thread_sibling(policy->cpu);
@@ -475,7 +688,34 @@
} else {
kernfs_put(kn);
}
- return cpufreq_table_validate_and_show(policy, powernv_freqs);
+
+ gpstates = kzalloc(sizeof(*gpstates), GFP_KERNEL);
+ if (!gpstates)
+ return -ENOMEM;
+
+ policy->driver_data = gpstates;
+
+ /* initialize timer */
+ init_timer_deferrable(&gpstates->timer);
+ gpstates->timer.data = (unsigned long)policy;
+ gpstates->timer.function = gpstate_timer_handler;
+ gpstates->timer.expires = jiffies +
+ msecs_to_jiffies(GPSTATE_TIMER_INTERVAL);
+ spin_lock_init(&gpstates->gpstate_lock);
+ ret = cpufreq_table_validate_and_show(policy, powernv_freqs);
+
+ if (ret < 0)
+ kfree(policy->driver_data);
+
+ return ret;
+}
+
+static int powernv_cpufreq_cpu_exit(struct cpufreq_policy *policy)
+{
+ /* timer is deleted in cpufreq_cpu_stop() */
+ kfree(policy->driver_data);
+
+ return 0;
}
static int powernv_cpufreq_reboot_notifier(struct notifier_block *nb,
@@ -603,15 +843,19 @@
static void powernv_cpufreq_stop_cpu(struct cpufreq_policy *policy)
{
struct powernv_smp_call_data freq_data;
+ struct global_pstate_info *gpstates = policy->driver_data;
freq_data.pstate_id = powernv_pstate_info.min;
+ freq_data.gpstate_id = powernv_pstate_info.min;
smp_call_function_single(policy->cpu, set_pstate, &freq_data, 1);
+ del_timer_sync(&gpstates->timer);
}
static struct cpufreq_driver powernv_cpufreq_driver = {
.name = "powernv-cpufreq",
.flags = CPUFREQ_CONST_LOOPS,
.init = powernv_cpufreq_cpu_init,
+ .exit = powernv_cpufreq_cpu_exit,
.verify = cpufreq_generic_frequency_table_verify,
.target_index = powernv_cpufreq_target_index,
.get = powernv_cpufreq_get,