blk-iocost: fix busy_level reset when no IOs complete
Subsystems (3)
PATCH v2 1/1 blk-iocost: fix busy_level reset when no IOs complete
When a disk is saturated, it is common for no IOs to complete within a timer period. Currently, in this case, rq_wait_pct and missed_ppm are calculated as 0, the iocost incorrectly interprets this as meeting QoS targets and resets busy_level to 0.
This reset prevents busy_level from reaching the threshold (4) needed to reduce vrate. On certain cloud storage, such as Azure Premium SSD, we observed that iocost may fail to reduce vrate for tens of seconds during saturation, failing to mitigate noisy neighbor issues.
Fix this by tracking the number of IO completions (nr_done) in a period. If nr_done is 0, we adjust the logic:
* If there are lagging IOs, the saturation status is unknown, so we try to keep busy_level unchanged. To avoid drastic vrate oscillations, we clamp it between -4 and 4.
* If there are shortages but no lagging IOs, the vrate might be too low to issue any IOs. We should allow vrate to increase but not decrease.
* Otherwise, reset busy_level to 0.
Note that when nr_done is 0 and nr_lagging is 0, the adjustment logic is nearly identical to the "QoS targets are being met with >25% margin" state, which minimizes the risk of regressions.
The issue is consistently reproducible on Azure Standard_D8as_v5 (Dasv5) VMs with 512GB Premium SSD (P20) using the script below. It was not observed on GCP n2d VMs (100G pd-ssd and 1.5T local-ssd), and no regressions were found with this patch. In this script, cgA saturates the device. The iocost is expected to throttle it so that cgB's completion latency remains low.
BLK_DEVID="8:0"
MODEL="rbps=173471131 rseqiops=3566 rrandiops=3566 wbps=173333269 wseqiops=3566 wrandiops=3566" QOS="rpct=90.00 rlat=3500 wpct=90 wlat=3500 min=80 max=10000"
echo "$BLK_DEVID ctrl=user model=linear $MODEL" > /sys/fs/cgroup/io.cost.model echo "$BLK_DEVID enable=1 ctrl=user $QOS" > /sys/fs/cgroup/io.cost.qos
CG_A="/sys/fs/cgroup/cgA"
CG_B="/sys/fs/cgroup/cgB"
FILE_A="/data0/A.fio.testfile" FILE_B="/data0/B.fio.testfile" RESULT_DIR="./iocost_results_$(date +%Y%m%d_%H%M%S)"
mkdir -p "$CG_A" "$CG_B" "$RESULT_DIR"
get_result() {
local file=$1
local label=$2 local results=$(jq -r '
.jobs[0].mixed |
( .iops | tonumber | round ) as $iops |
( .bw_bytes / 1024 / 1024 ) as $bps |
( .clat_ns.mean / 1000000 ) as $avg |
( .clat_ns.max / 1000000 ) as $max |
( .clat_ns.percentile["90.000000"] / 1000000 ) as $p90 |
( .clat_ns.percentile["99.000000"] / 1000000 ) as $p99 |
( .clat_ns.percentile["99.900000"] / 1000000 ) as $p999 |
( .clat_ns.percentile["99.990000"] / 1000000 ) as $p9999 |
"\($iops)|\($bps)|\($avg)|\($max)|\($p90)|\($p99)|\($p999)|\($p9999)"
' "$file")
IFS='|' read -r iops bps avg max p90 p99 p999 p9999 <<<"$results"
printf "%-8s %-6s %-7.2f %-8.2f %-8.2f %-8.2f %-8.2f %-8.2f %-8.2f\n" \
"$label" "$iops" "$bps" "$avg" "$max" "$p90" "$p99" "$p999" "$p9999"}
run_fio() {
local cg_path=$1
local filename=$2
local name=$3
local bs=$4
local qd=$5
local out=$6
shift 6
local extra=$@ (
pid=$(sh -c 'echo $PPID')
echo $pid >"${cg_path}/cgroup.procs"
fio --name="$name" --filename="$filename" --direct=1 --rw=randrw --rwmixread=50 \
--ioengine=libaio --bs="$bs" --iodepth="$qd" --size=4G --runtime=10 \
--time_based --group_reporting --unified_rw_reporting=mixed \
--output-format=json --output="$out" $extra >/dev/null 2>&1
) &}
echo "Starting Test ..."
for bs_b in "4k" "32k" "256k"; do
echo "Running iteration: BS=$bs_b"
out_a="${RESULT_DIR}/cgA_1m.json"
out_b="${RESULT_DIR}/cgB_${bs_b}.json" # cgA: Heavy background (BS 1MB, QD 128)
run_fio "$CG_A" "$FILE_A" "cgA" "1m" 128 "$out_a"
# cgB: Latency sensitive (Variable BS, QD 1, Read/Write IOPS limit 100)
run_fio "$CG_B" "$FILE_B" "cgB" "$bs_b" 1 "$out_b" "--rate_iops=100"
wait
SUMMARY_DATA+="$(get_result "$out_a" "cgA-1m")"$'\n'
SUMMARY_DATA+="$(get_result "$out_b" "cgB-$bs_b")"$'\n\n'done
# Final Output
echo -e "\nFinal Results Summary:\n"
printf "%-8s %-6s %-7s %-8s %-8s %-8s %-8s %-8s %-8s\n\n" \
"CGROUP" "IOPS" "MB/s" "Avg(ms)" "Max(ms)" "P90(ms)" "P99" "P99.9" "P99.99"
echo "$SUMMARY_DATA"echo "Results saved in $RESULT_DIR"
Before:
CGROUP IOPS MB/s Avg(ms) Max(ms) P90(ms) P99 P99.9 P99.99
cgA-1m 167 167.02 748.65 1641.43 960.50 1551.89 1635.78 1635.78 cgB-4k 5 0.02 190.57 806.84 742.39 809.50 809.50 809.50
cgA-1m 166 166.36 751.38 1744.31 994.05 1451.23 1736.44 1736.44 cgB-32k 4 0.14 225.71 1057.25 759.17 1061.16 1061.16 1061.16
cgA-1m 166 165.91 751.48 1610.94 1010.83 1417.67 1602.22 1619.00 cgB-256k 5 1.26 198.50 1046.30 742.39 1044.38 1044.38 1044.38
After:
CGROUP IOPS MB/s Avg(ms) Max(ms) P90(ms) P99 P99.9 P99.99
cgA-1m 159 158.59 769.06 828.52 809.50 817.89 826.28 826.28 cgB-4k 200 0.78 2.01 26.11 2.87 6.26 12.39 26.08
cgA-1m 147 146.84 832.05 985.80 943.72 960.50 985.66 985.66 cgB-32k 200 6.25 2.82 71.05 3.42 15.40 50.07 70.78
cgA-1m 114 114.47 1044.98 1294.48 1199.57 1283.46 1300.23 1300.23 cgB-256k 200 50.00 4.01 34.49 5.08 15.66 30.54 34.34
---
v2:
- Handle more edge cases to prevent potential regressions.
v1: https://lore.kernel.org/all/[email protected]/
block/blk-iocost.c | 35 +++++++++++++++++++++++++++++------ 1 file changed, 29 insertions(+), 6 deletions(-)
block/blk-iocost.c +29 -6
@@ -1596,7 +1596,8 @@ static enum hrtimer_restart iocg_waitq_timer_fn(struct hrtimer *timer)return HRTIMER_NORESTART ; } -static void ioc_lat_stat (struct ioc *ioc,u32 *missed_ppm_ar,u32 *rq_wait_pct_p) +static void ioc_lat_stat (struct ioc *ioc,u32 *missed_ppm_ar,u32 *rq_wait_pct_p, +u32 *nr_done) {u32 nr_met[2 ] = { };u32 nr_missed[2 ] = { };
@@ -1633,6 +1634,8 @@ static void ioc_lat_stat(struct ioc *ioc, u32 *missed_ppm_ar, u32 *rq_wait_pct_p *rq_wait_pct_p =div64_u64 (rq_wait_ns *100 , ioc->period_us *NSEC_PER_USEC ); + + *nr_done = nr_met[READ ] + nr_met[WRITE ] + nr_missed[READ ] + nr_missed[WRITE ]; }/* was iocg idle this period? */
@@ -2250,12 +2253,12 @@ static void ioc_timer_fn(struct timer_list *timer)u64 usage_us_sum =0 ;u32 ppm_rthr;u32 ppm_wthr; -u32 missed_ppm[2 ], rq_wait_pct; +u32 missed_ppm[2 ], rq_wait_pct, nr_done;u64 period_vtime;int prev_busy_level;/* how were the latencies during the period? */ -ioc_lat_stat (ioc, missed_ppm, &rq_wait_pct); +ioc_lat_stat (ioc, missed_ppm, &rq_wait_pct, &nr_done);/* take care of active iocgs */ spin_lock_irq (&ioc->lock);
@@ -2397,9 +2400,29 @@ static void ioc_timer_fn(struct timer_list *timer)* and should increase vtime rate. */ prev_busy_level = ioc->busy_level; -if (rq_wait_pct >RQ_WAIT_BUSY_PCT || - missed_ppm[READ ] > ppm_rthr || - missed_ppm[WRITE ] > ppm_wthr) { +if (!nr_done) { +if (nr_lagging) +/* +* When there are lagging IOs but no completions, we +* don't know if the IO latency will meet the QoS +* targets. The disk might be saturated or not. We +* should not reset busy_level to 0 (which would +* prevent vrate from scaling up or down), but rather +* try to keep it unchanged. To avoid drastic vrate +* oscillations, we clamp it between -4 and 4. +*/ + ioc->busy_level =clamp (ioc->busy_level,-4 ,4 ); +else if (nr_shortages) +/* +* The vrate might be too low to issue any IOs. We +* should allow vrate to increase but not decrease. +*/ + ioc->busy_level =min (ioc->busy_level,0 ); +else + ioc->busy_level =0 ; + }else if (rq_wait_pct >RQ_WAIT_BUSY_PCT || + missed_ppm[READ ] > ppm_rthr || + missed_ppm[WRITE ] > ppm_wthr) {/* clearly missing QoS targets, slow down vrate */ ioc->busy_level =max (ioc->busy_level,0 ); ioc->busy_level++;
Signature · 1 line
2.53.0
Tejun Heo
maintainerHello,
On Sun, Mar 29, 2026 at 03:41:12PM +0000, Jialin Wang wrote:
...
Quoted 15 lines
Before:
CGROUP IOPS MB/s Avg(ms) Max(ms) P90(ms) P99 P99.9 P99.99cgA-1m 167 167.02 748.65 1641.43 960.50 1551.89 1635.78 1635.78 cgB-4k 5 0.02 190.57 806.84 742.39 809.50 809.50 809.50cgA-1m 166 166.36 751.38 1744.31 994.05 1451.23 1736.44 1736.44 cgB-32k 4 0.14 225.71 1057.25 759.17 1061.16 1061.16 1061.16cgA-1m 166 165.91 751.48 1610.94 1010.83 1417.67 1602.22 1619.00 cgB-256k 5 1.26 198.50 1046.30 742.39 1044.38 1044.38 1044.38After:
CGROUP IOPS MB/s Avg(ms) Max(ms) P90(ms) P99 P99.9 P99.99
Quoted 8 lines
cgA-1m 159 158.59 769.06 828.52 809.50 817.89 826.28 826.28 cgB-4k 200 0.78 2.01 26.11 2.87 6.26 12.39 26.08cgA-1m 147 146.84 832.05 985.80 943.72 960.50 985.66 985.66 cgB-32k 200 6.25 2.82 71.05 3.42 15.40 50.07 70.78cgA-1m 114 114.47 1044.98 1294.48 1199.57 1283.46 1300.23 1300.23 cgB-256k 200 50.00 4.01 34.49 5.08 15.66 30.54 34.34
Are the latency numbers end-to-end or on-device? If former, can you provide on-device numbers? What period duration are you using?
block/blk-iocost.c static void ioc_timer_fn(struct timer_list *timer) · line 2404
Quoted 9 lines
block/blk-iocost.c +2
-3@@ -2397,9 +2400,29 @@ static void ioc_timer_fn(struct timer_list *timer)* and should increase vtime rate. */ prev_busy_level = ioc->busy_level; -if (rq_wait_pct >RQ_WAIT_BUSY_PCT || - missed_ppm[READ ] > ppm_rthr || - missed_ppm[WRITE ] > ppm_wthr) { +if (!nr_done) { +if (nr_lagging)
Please use {} even when it's just comments that makes the bodies multi-line.
block/blk-iocost.c static void ioc_timer_fn(struct timer_list *timer) · line 2414
Quoted 10 lines
block/blk-iocost.c +10
-0+/* +* When there are lagging IOs but no completions, we +* don't know if the IO latency will meet the QoS +* targets. The disk might be saturated or not. We +* should not reset busy_level to 0 (which would +* prevent vrate from scaling up or down), but rather +* try to keep it unchanged. To avoid drastic vrate +* oscillations, we clamp it between -4 and 4. +*/ + ioc->busy_level =clamp (ioc->busy_level,-4 ,4 );
Is this from some observed behavior or just out of intuition? The justification seems a bit flimsy. Why -4 and 4?
block/blk-iocost.c static void ioc_timer_fn(struct timer_list *timer) · line 2420
Quoted 6 lines
block/blk-iocost.c +6
-0+else if (nr_shortages) +/* +* The vrate might be too low to issue any IOs. We +* should allow vrate to increase but not decrease. +*/ + ioc->busy_level =min (ioc->busy_level,0 );
So, this is no completion, no lagging and shortages case. In the existing code, this would alos get busy_level-- to get things moving. Wouldn't this path need that too? Or rather, would it make more sense to handle !nr_done && nr_lagging case and leave the other cases as-are?
Thanks.
Signature · 1 line
tejun
Hi,
Quoted 22 lines · Tejun Heo
Hello,
On Sun, Mar 29, 2026 at 03:41:12PM +0000, Jialin Wang wrote:
...Quoted 18 lines
Before:
CGROUP IOPS MB/s Avg(ms) Max(ms) P90(ms) P99 P99.9 P99.99cgA-1m 167 167.02 748.65 1641.43 960.50 1551.89 1635.78 1635.78 cgB-4k 5 0.02 190.57 806.84 742.39 809.50 809.50 809.50cgA-1m 166 166.36 751.38 1744.31 994.05 1451.23 1736.44 1736.44 cgB-32k 4 0.14 225.71 1057.25 759.17 1061.16 1061.16 1061.16cgA-1m 166 165.91 751.48 1610.94 1010.83 1417.67 1602.22 1619.00 cgB-256k 5 1.26 198.50 1046.30 742.39 1044.38 1044.38 1044.38After:
CGROUP IOPS MB/s Avg(ms) Max(ms) P90(ms) P99 P99.9 P99.99cgA-1m 159 158.59 769.06 828.52 809.50 817.89 826.28 826.28 cgB-4k 200 0.78 2.01 26.11 2.87 6.26 12.39 26.08
Quoted 8 lines
Quoted 5 lines
cgA-1m 147 146.84 832.05 985.80 943.72 960.50 985.66 985.66 cgB-32k 200 6.25 2.82 71.05 3.42 15.40 50.07 70.78cgA-1m 114 114.47 1044.98 1294.48 1199.57 1283.46 1300.23 1300.23 cgB-256k 200 50.00 4.01 34.49 5.08 15.66 30.54 34.34Are the latency numbers end-to-end or on-device? If former, can you provide on-device numbers? What period duration are you using?
These latency numbers are completion latency results from fio using ioengine=libaio. For cgB, since --iodepth=1 is used, these completion latencies are very close to the actual on-device times.
I used the following QoS parameters:
rpct=90 rlat=3500 wpct=90 wlat=3500 min=80 max=10000 (period: 7ms)
When switching to:
rpct=80 rlat=10000 wpct=80 wlat=10000 min=80 max=10000 (period: 40ms) While this showed some improvement, cgB still failed to reach the expected 200 IOPS, and the P99 latency remained high:
CGROUP IOPS MB/s Avg(ms) Max(ms) P90(ms) P99 P99.9 P99.99
cgA-1m 161 160.81 758.52 1462.38 1044.38 1317.01 1451.23 1468.01 cgB-4k 125 0.49 7.18 661.39 2.70 189.79 650.12 658.51
cgA-1m 155 154.63 784.92 1234.01 1010.83 1182.79 1233.13 1233.13 cgB-32k 136 4.26 6.40 300.78 3.85 160.43 295.70 299.89
cgA-1m 138 137.91 860.32 1704.14 1317.01 1669.33 1702.89 1702.89 cgB-256k 95 23.70 9.83 394.73 5.34 206.57 396.36 396.36
I also tested several other sets of parameters and the results were similar. Using bpftrace, it can still be frequently observed that busy_level is reset to 0 when no IO complete, and the vrate cannot be lowered in time.
08:26:20.186950 iocost_ioc_vrate_adj: [sdb] vrate=127.50%->126.23% busy=4 missed_ppm=1000000:1000000 rq_wait_pct=0 lagging=3 shortages=1 08:26:20.220910 ioc_rqos_done
08:26:20.222616 ioc_rqos_done
08:26:20.226913 ioc_rqos_done
08:26:20.227951 iocost_ioc_vrate_adj: [sdb] vrate=126.23%->124.97% busy=5 missed_ppm=1000000:1000000 rq_wait_pct=0 lagging=3 shortages=1
-- no IO complete, busy_level was reset to 0 --
08:26:20.268945 iocost_ioc_vrate_adj: [sdb] vrate=124.97%->124.97% busy=0 missed_ppm=0:0 rq_wait_pct=0 lagging=3 shortages=1
bpftrace -e '
#define VTIME_PER_USEC 137438
kfunc:ioc_rqos_done
{
printf("%s ioc_rqos_done\n", strftime("%H:%M:%S.%f", nsecs));
}tracepoint:iocost:iocost_ioc_vrate_adj
{
$old_vrate = args->old_vrate * 10000 / VTIME_PER_USEC;
$new_vrate = args->new_vrate * 10000 / VTIME_PER_USEC;
printf("%s iocost_ioc_vrate_adj: [%s] vrate=%d.%02d%%->%d.%02d%% busy=%d missed_ppm=%u:%u rq_wait_pct=%u lagging=%d shortages=%d\n",
strftime("%H:%M:%S.%f", nsecs), str(args->devname),
$old_vrate / 100, $old_vrate % 100, $new_vrate / 100,
$new_vrate % 100, args->busy_level, args->read_missed_ppm,
args->write_missed_ppm, args->rq_wait_pct, args->nr_lagging,
args->nr_shortages);
}'Quoted 12 lines
block/blk-iocost.c static void ioc_timer_fn(struct timer_list *timer) · line 2404
Quoted 9 lines
block/blk-iocost.c +2
-3@@ -2397,9 +2400,29 @@ static void ioc_timer_fn(struct timer_list *timer)* and should increase vtime rate. */ prev_busy_level = ioc->busy_level; -if (rq_wait_pct >RQ_WAIT_BUSY_PCT || - missed_ppm[READ ] > ppm_rthr || - missed_ppm[WRITE ] > ppm_wthr) { +if (!nr_done) { +if (nr_lagging)Please use {} even when it's just comments that makes the bodies multi-line.
Quoted 13 lines
block/blk-iocost.c static void ioc_timer_fn(struct timer_list *timer) · line 2414
Quoted 10 lines
block/blk-iocost.c +10
-0+/* +* When there are lagging IOs but no completions, we +* don't know if the IO latency will meet the QoS +* targets. The disk might be saturated or not. We +* should not reset busy_level to 0 (which would +* prevent vrate from scaling up or down), but rather +* try to keep it unchanged. To avoid drastic vrate +* oscillations, we clamp it between -4 and 4. +*/ + ioc->busy_level =clamp (ioc->busy_level,-4 ,4 );Is this from some observed behavior or just out of intuition? The justification seems a bit flimsy. Why -4 and 4?
During my testing with the parameters rpct=90 rlat=3500 wpct=90 wlat=3500 min=10 max=10000, I noticed that vrate occasionally drops significantly (down to 50% or lower), which adversely impacted the IOPS of cgA. So I limit the busy_level to a maximum of 4 to reduce vrate at the lowest speed.
CGROUP IOPS MB/s Avg(ms) Max(ms) P90(ms) P99 P99.9 P99.99 cgA-1m 137 137.11 891.21 1278.66 1082.13 1216.35 1266.68 1283.46 cgB-4k 200 0.78 2.12 62.64 2.47 7.44 49.55 62.65
I realized that raising min to 80 would effectively mitigate this issue, so I will remove it in the next v3.
Quoted 11 lines
block/blk-iocost.c static void ioc_timer_fn(struct timer_list *timer) · line 2420
Quoted 6 lines
block/blk-iocost.c +6
-0+else if (nr_shortages) +/* +* The vrate might be too low to issue any IOs. We +* should allow vrate to increase but not decrease. +*/ + ioc->busy_level =min (ioc->busy_level,0 );So, this is no completion, no lagging and shortages case. In the existing code, this would alos get busy_level-- to get things moving. Wouldn't this path need that too? Or rather, would it make more sense to handle !nr_done &&
nr_laggingcase and leave the other cases as-are?
That's a fair point. My initial thought was not to adjust busy_level when there is no latency data, and I haven't observed this specific path (no completions, no lagging, but with shortages) occurring in my testing so far, so I might have been overthinking it. I will simplify the logic in v3 to handle only the !nr_done && nr_lagging case and leave the other cases as they are.
Signature · 2 lines
Thanks, Jialin
PATCH v3 1/1 blk-iocost: fix busy_level reset when no IOs complete
In reply to Jialin WangWhen a disk is saturated, it is common for no IOs to complete within a timer period. Currently, in this case, rq_wait_pct and missed_ppm are calculated as 0, the iocost incorrectly interprets this as meeting QoS targets and resets busy_level to 0.
This reset prevents busy_level from reaching the threshold (4) needed to reduce vrate. On certain cloud storage, such as Azure Premium SSD, we observed that iocost may fail to reduce vrate for tens of seconds during saturation, failing to mitigate noisy neighbor issues.
Fix this by tracking the number of IO completions (nr_done) in a period. If nr_done is 0 and there are lagging IOs, the saturation status is unknown, so we keep busy_level unchanged.
The issue is consistently reproducible on Azure Standard_D8as_v5 (Dasv5) VMs with 512GB Premium SSD (P20) using the script below. It was not observed on GCP n2d VMs (with 100G pd-ssd and 1.5T local-ssd), and no regressions were found with this patch. In this script, cgA performs large IOs with iodepth=128, while cgB performs small IOs with iodepth=1 rate_iops=100 rw=randrw. With iocost enabled, we expect it to throttle cgA, the submission latency (slat) of cgA should be significantly higher, cgB can reach 200 IOPS and the completion latency (clat) should below.
BLK_DEVID="8:0"
MODEL="rbps=173471131 rseqiops=3566 rrandiops=3566 wbps=173333269 wseqiops=3566 wrandiops=3566" QOS="rpct=90 rlat=3500 wpct=90 wlat=3500 min=80 max=10000"
echo "$BLK_DEVID ctrl=user model=linear $MODEL" > /sys/fs/cgroup/io.cost.model echo "$BLK_DEVID enable=1 ctrl=user $QOS" > /sys/fs/cgroup/io.cost.qos
CG_A="/sys/fs/cgroup/cgA"
CG_B="/sys/fs/cgroup/cgB"
FILE_A="/path/to/sda/A.fio.testfile" FILE_B="/path/to/sda/B.fio.testfile" RESULT_DIR="./iocost_results_$(date +%Y%m%d_%H%M%S)"
mkdir -p "$CG_A" "$CG_B" "$RESULT_DIR"
get_result() {
local file=$1
local label=$2 local results=$(jq -r '
.jobs[0].mixed |
( .iops | tonumber | round ) as $iops |
( .bw_bytes / 1024 / 1024 ) as $bps |
( .slat_ns.mean / 1000000 ) as $slat |
( .clat_ns.mean / 1000000 ) as $avg |
( .clat_ns.max / 1000000 ) as $max |
( .clat_ns.percentile["90.000000"] / 1000000 ) as $p90 |
( .clat_ns.percentile["99.000000"] / 1000000 ) as $p99 |
( .clat_ns.percentile["99.900000"] / 1000000 ) as $p999 |
( .clat_ns.percentile["99.990000"] / 1000000 ) as $p9999 |
"\($iops)|\($bps)|\($slat)|\($avg)|\($max)|\($p90)|\($p99)|\($p999)|\($p9999)"
' "$file")
IFS='|' read -r iops bps slat avg max p90 p99 p999 p9999 <<<"$results"
printf "%-8s %-6s %-7.2f %-8.2f %-8.2f %-8.2f %-8.2f %-8.2f %-8.2f %-8.2f\n" \
"$label" "$iops" "$bps" "$slat" "$avg" "$max" "$p90" "$p99" "$p999" "$p9999"}
run_fio() {
local cg_path=$1
local filename=$2
local name=$3
local bs=$4
local qd=$5
local out=$6
shift 6
local extra=$@ (
pid=$(sh -c 'echo $PPID')
echo $pid >"${cg_path}/cgroup.procs"
fio --name="$name" --filename="$filename" --direct=1 --rw=randrw --rwmixread=50 \
--ioengine=libaio --bs="$bs" --iodepth="$qd" --size=4G --runtime=10 \
--time_based --group_reporting --unified_rw_reporting=mixed \
--output-format=json --output="$out" $extra >/dev/null 2>&1
) &}
echo "Starting Test ..."
for bs_b in "4k" "32k" "256k"; do
echo "Running iteration: BS=$bs_b"
out_a="${RESULT_DIR}/cgA_1m.json"
out_b="${RESULT_DIR}/cgB_${bs_b}.json" # cgA: Heavy background (BS 1MB, QD 128)
run_fio "$CG_A" "$FILE_A" "cgA" "1m" 128 "$out_a"
# cgB: Latency sensitive (Variable BS, QD 1, Read/Write IOPS limit 100)
run_fio "$CG_B" "$FILE_B" "cgB" "$bs_b" 1 "$out_b" "--rate_iops=100"
wait
SUMMARY_DATA+="$(get_result "$out_a" "cgA-1m")"$'\n'
SUMMARY_DATA+="$(get_result "$out_b" "cgB-$bs_b")"$'\n\n'done
echo -e "\nFinal Results Summary:\n"
printf "%-8s %-6s %-7s %-8s %-8s %-8s %-8s %-8s %-8s %-8s\n" \
"" "" "" "slat" "clat" "clat" "clat" "clat" "clat" "clat"
printf "%-8s %-6s %-7s %-8s %-8s %-8s %-8s %-8s %-8s %-8s\n\n" \
"CGROUP" "IOPS" "MB/s" "avg(ms)" "avg(ms)" "max(ms)" "P90(ms)" "P99" "P99.9" "P99.99"
echo "$SUMMARY_DATA"echo "Results saved in $RESULT_DIR"
Before:
slat clat clat clat clat clat clat CGROUP IOPS MB/s avg(ms) avg(ms) max(ms) P90(ms) P99 P99.9 P99.99
cgA-1m 166 166.37 3.44 748.95 1298.29 977.27 1233.13 1300.23 1300.23 cgB-4k 5 0.02 0.02 181.74 761.32 742.39 759.17 759.17 759.17
cgA-1m 167 166.51 1.98 748.68 1549.41 809.50 1451.23 1551.89 1551.89 cgB-32k 6 0.18 0.02 169.98 761.76 742.39 759.17 759.17 759.17
cgA-1m 166 165.55 2.89 750.89 1540.37 851.44 1451.23 1535.12 1535.12 cgB-256k 5 1.30 0.02 191.35 759.51 750.78 759.17 759.17 759.17
After:
slat clat clat clat clat clat clat CGROUP IOPS MB/s avg(ms) avg(ms) max(ms) P90(ms) P99 P99.9 P99.99
cgA-1m 162 162.48 6.14 749.69 850.02 826.28 834.67 843.06 851.44 cgB-4k 199 0.78 0.01 1.95 42.12 2.57 7.50 34.87 42.21
cgA-1m 146 146.20 6.83 833.04 908.68 893.39 901.78 910.16 910.16 cgB-32k 200 6.25 0.01 2.32 31.40 3.06 7.50 16.58 31.33
cgA-1m 110 110.46 9.04 1082.67 1197.91 1182.79 1199.57 1199.57 1199.57 cgB-256k 200 49.98 0.02 3.69 22.20 4.88 9.11 20.05 22.15
---
Changes in v3:
- Handle only the !nr_done && nr_lagging case and leave the other cases as they are.
Changes in v2:
- Handle more edge cases to prevent potential regressions.
v2: https://lore.kernel.org/all/[email protected]/
v1: https://lore.kernel.org/all/[email protected]/
block/blk-iocost.c | 23 +++++++++++++++++------ 1 file changed, 17 insertions(+), 6 deletions(-)
block/blk-iocost.c +17 -6
@@ -1596,7 +1596,8 @@ static enum hrtimer_restart iocg_waitq_timer_fn(struct hrtimer *timer)return HRTIMER_NORESTART ; } -static void ioc_lat_stat (struct ioc *ioc,u32 *missed_ppm_ar,u32 *rq_wait_pct_p) +static void ioc_lat_stat (struct ioc *ioc,u32 *missed_ppm_ar,u32 *rq_wait_pct_p, +u32 *nr_done) {u32 nr_met[2 ] = { };u32 nr_missed[2 ] = { };
@@ -1633,6 +1634,8 @@ static void ioc_lat_stat(struct ioc *ioc, u32 *missed_ppm_ar, u32 *rq_wait_pct_p *rq_wait_pct_p =div64_u64 (rq_wait_ns *100 , ioc->period_us *NSEC_PER_USEC ); + + *nr_done = nr_met[READ ] + nr_met[WRITE ] + nr_missed[READ ] + nr_missed[WRITE ]; }/* was iocg idle this period? */
@@ -2250,12 +2253,12 @@ static void ioc_timer_fn(struct timer_list *timer)u64 usage_us_sum =0 ;u32 ppm_rthr;u32 ppm_wthr; -u32 missed_ppm[2 ], rq_wait_pct; +u32 missed_ppm[2 ], rq_wait_pct, nr_done;u64 period_vtime;int prev_busy_level;/* how were the latencies during the period? */ -ioc_lat_stat (ioc, missed_ppm, &rq_wait_pct); +ioc_lat_stat (ioc, missed_ppm, &rq_wait_pct, &nr_done);/* take care of active iocgs */ spin_lock_irq (&ioc->lock);
@@ -2397,9 +2400,17 @@ static void ioc_timer_fn(struct timer_list *timer)* and should increase vtime rate. */ prev_busy_level = ioc->busy_level; -if (rq_wait_pct >RQ_WAIT_BUSY_PCT || - missed_ppm[READ ] > ppm_rthr || - missed_ppm[WRITE ] > ppm_wthr) { +if (!nr_done && nr_lagging) { +/* +* When there are lagging IOs but no completions, we don't +* know if the IO latency will meet the QoS targets. The +* disk might be saturated or not. We should not reset +* busy_level to 0 (which would prevent vrate from scaling +* up or down), but rather to keep it unchanged. +*/ + }else if (rq_wait_pct >RQ_WAIT_BUSY_PCT || + missed_ppm[READ ] > ppm_rthr || + missed_ppm[WRITE ] > ppm_wthr) {/* clearly missing QoS targets, slow down vrate */ ioc->busy_level =max (ioc->busy_level,0 ); ioc->busy_level++;
Signature · 1 line
2.53.0
Quoted 22 lines · Jialin Wang
When a disk is saturated, it is common for no IOs to complete within a timer period. Currently, in this case,
rq_wait_pctandmissed_ppmare calculated as 0, the iocost incorrectly interprets this as meeting QoS targets and resetsbusy_levelto 0.This reset prevents
busy_levelfrom reaching the threshold (4) needed to reduce vrate. On certain cloud storage, such as Azure Premium SSD, we observed that iocost may fail to reduce vrate for tens of seconds during saturation, failing to mitigate noisy neighbor issues.Fix this by tracking the number of IO completions (
nr_done) in a period. Ifnr_doneis 0 and there are lagging IOs, the saturation status is unknown, so we keepbusy_levelunchanged.The issue is consistently reproducible on Azure
Standard_D8as_v5(Dasv5) VMs with 512GB Premium SSD (P20) using the script below. It was not observed on GCP n2d VMs (with 100G pd-ssd and 1.5T local-ssd), and no regressions were found with this patch. In this script, cgA performs large IOs with iodepth=128, while cgB performs small IOs with iodepth=1 rate_iops=100 rw=randrw. With iocost enabled, we expect it to throttle cgA, the submission latency (slat) of cgA should be significantly higher, cgB can reach 200 IOPS and the completion latency (clat) should below.
...
Quoted 1 line
Signed-off-by Jialin Wang <[email protected]>
Thanks.
Signature · 1 line
tejun
Quoted 11 lines · Jialin Wang
When a disk is saturated, it is common for no IOs to complete within a timer period. Currently, in this case,
rq_wait_pctandmissed_ppmare calculated as 0, the iocost incorrectly interprets this as meeting QoS targets and resetsbusy_levelto 0.This reset prevents
busy_levelfrom reaching the threshold (4) needed to reduce vrate. On certain cloud storage, such as Azure Premium SSD, we observed that iocost may fail to reduce vrate for tens of seconds during saturation, failing to mitigate noisy neighbor issues.[...]
Applied, thanks!
[1/1] blk-iocost: fix busy_level reset when no IOs complete
commit: f91ffe89b2016d280995a9c28d73288b02d83615
Best regards,
Signature · 1 line
Jens Axboe