In my production environment i had a 8 core, around 40GB+ RAM server. I am seeing my docker containers were using more cpu cycles, and making other standalone deployed services slow.
Ex: we had a nifi service running in a container using cpu around 300%(varying one) and other containers like database, kafka some times.
here is the docker container inspect i found
"Isolation": "",
"CpuShares": 0,
"Memory": 0,
"NanoCpus": 0,
"CgroupParent": "",
"BlkioWeight": 0,
"BlkioWeightDevice": null,
"BlkioDeviceReadBps": null,
"BlkioDeviceWriteBps": null,
"BlkioDeviceReadIOps": null,
"BlkioDeviceWriteIOps": null,
"CpuPeriod": 0,
"CpuQuota": 0,
"CpuRealtimePeriod": 0,
"CpuRealtimeRuntime": 0,
"CpusetCpus": "",
"CpusetMems": "",
"Devices": null,
"DeviceCgroupRules": null,
"DiskQuota": 0,
"KernelMemory": 0,
"MemoryReservation": 0,
"MemorySwap": 0,
"MemorySwappiness": null,
"OomKillDisable": false,
"PidsLimit": 0,
"Ulimits": null,
"CpuCount": 0,
"CpuPercent": 0,
"IOMaximumIOps": 0,
"IOMaximumBandwidth": 0
here is the top usage
$top
top - 01:08:47 up 15 days, 3:23, 1 user, load average: 5.73, 5.58, 5.44
Tasks: 320 total, 4 running, 314 sleeping, 0 stopped, 2 zombie
%Cpu0 : 15.9 us, 6.2 sy, 0.0 ni, 77.2 id, 0.0 wa, 0.0 hi, 0.7 si, 0.0 st
%Cpu1 : 26.8 us, 3.1 sy, 0.0 ni, 69.4 id, 0.0 wa, 0.0 hi, 0.7 si, 0.0 st
%Cpu2 : 8.9 us, 5.2 sy, 0.0 ni, 84.9 id, 0.0 wa, 0.0 hi, 1.0 si, 0.0 st
%Cpu3 : 13.8 us, 6.0 sy, 0.0 ni, 78.8 id, 0.0 wa, 0.0 hi, 1.4 si, 0.0 st
%Cpu4 : 34.4 us, 3.8 sy, 0.0 ni, 61.1 id, 0.3 wa, 0.0 hi, 0.3 si, 0.0 st
%Cpu5 : 81.3 us, 4.9 sy, 0.0 ni, 11.3 id, 0.0 wa, 0.0 hi, 2.5 si, 0.0 st
%Cpu6 : 64.8 us, 2.5 sy, 0.0 ni, 31.7 id, 0.0 wa, 0.0 hi, 1.1 si, 0.0 st
%Cpu7 : 77.4 us, 5.0 sy, 0.0 ni, 16.5 id, 0.0 wa, 0.0 hi, 1.1 si, 0.0 st
KiB Mem: 57803616 total, 51826940 used, 5976676 free, 1047628 buffers
KiB Swap: 16773116 total, 77040 used, 16696076 free. 19288708 cached Mem
in general when we services running directly on server like web application which takes some amount of cpu and memory. will docker containers completely use the cpu cycle making other services slow?
if yes, then what would be the best process to make both services use resources properly by not making system heavily loaded?
Thanks in advance !!!
Related
I am checking the impact of Linux's sched_rt_runtime_us.
My understanding of the Linux RT scheduling is sched_rt_period_us defines scheduling period of RT process, and sched_rt_runtime_us defines how much the RT process can run within that period.
In my Linux-4.18.20, the kernel.sched_rt_period_us = 1000000, kernel.sched_rt_runtime_us = 950000, so in each second, 95% time is used by RT process, 5% is for SCHED_OTHER processes.
By changing the kernel.sched_rt_runtime_us, the CPU usage of RT process shown in top should be proportional with sched_rt_runtime_us/sched_rt_period_us.
But my testing does NOT get the expected results, and what I got is as follows,
%CPU
kernel.sched_rt_runtime_us = 50000
2564 root rt 0 4516 748 684 R 19.9 0.0 0:37.82 testsched_top
kernel.sched_rt_runtime_us = 100000
2564 root rt 0 4516 748 684 R 40.5 0.0 0:23.16 testsched_top
kernel.sched_rt_runtime_us = 150000
2564 root rt 0 4516 748 684 R 60.1 0.0 0:53.29 testsched_top
kernel.sched_rt_runtime_us = 200000
2564 root rt 0 4516 748 684 R 80.1 0.0 1:24.96 testsched_top
The testsched_top is a SCHED_FIFO process with priority 99, and it is running in an isolated CPU.
The cgroup is configured in grub.cfg as cgroup_disable=cpuset,cpu,cpuacct to disable CPU related stuff.
I don't know why this happens, is there anything missing or wrong in my testing and understanding of Linux SCHED_FIFO scheduling?
N.B.: I am running this in Ubuntu VM, which is configured with 8 vCPUs, in which 4-7 are isolated to run RT processes. The host is Intel X86_64 with 6Cores (12 Threads), and there is NO other VMs running in the host. The above testsched_top was copied from https://viviendolared.blogspot.com/2017/03/death-by-real-time-scheduling.html?m=0, it sets priority 99 for SCHED_FIFO and loops indefinitely in one isolated CPU. I checked that isolated CPU usage, and got above results. –
I think I got the answer, and thank Rachid for the question.
In short, the kernel.sched_rt_period_us is the sum of RT time slice in a group of CPUs.
For example, in my 8vCPU VM configuration, CPU4-7 are isolated for running specific processes. So the kernel.sched_rt_period_us should be evenly divided among these 4 isolated CPUs, which means kernel.sched_rt_period_us/4 = 250000 is 100% CPU quota for each CPU in the isolated group. Setting kernel.sched_rt_period_us to 250000 makes the SCHED_FIFO process take all of the CPU. Accordingly, 25000 means 10% CPU usage for the CPU, 50000 means 20%, etc.
This is validated when CPU6 and CPU7 are isolated, in this case, 500000 can make the CPU to be 100% used by SCHED_FIFO process, 250000 makes 50% CPU usage.
Since these two kernel parameters are global ones, which means if the SCHED_FIFO process is put into the CPU0-5, 1000000/6 = 166000 should be the 100% quota for each CPU, 83000 makes 50% CPU usage, I also validated this.
Here is the snapshot of top,
%Cpu4 : 49.7 us, 0.0 sy, 0.0 ni, 50.3 id, 0.0 wa, 0.0 hi, 0.0 si, 0.0 st
%Cpu5 : 0.0 us, 0.0 sy, 0.0 ni,100.0 id, 0.0 wa, 0.0 hi, 0.0 si, 0.0 st
%Cpu6 : 0.0 us, 0.0 sy, 0.0 ni,100.0 id, 0.0 wa, 0.0 hi, 0.0 si, 0.0 st
%Cpu7 : 0.0 us, 0.0 sy, 0.0 ni,100.0 id, 0.0 wa, 0.0 hi, 0.0 si, 0.0 st
KiB Mem : 16422956 total, 14630144 free, 964880 used, 827932 buff/cache
KiB Swap: 1557568 total, 1557568 free, 0 used. 15245156 avail Mem
PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND
3748 root rt 0 4516 764 700 R 49.5 0.0 30:21.03 testsched_top
I have a question about Apache Spark. I set up an Apache Spark standalone cluster on my Ubuntu desktop. Then I wrote two lines in the spark_env.sh file: SPARK_WORKER_INSTANCES=4 and SPARK_WORKER_CORES=1. (I found that export is not necessary in spark_env.sh file if I start the cluster after I edit the spark_env.sh file.)
I wanted to have 4 worker instances in my single desktop and let them occupy 1 CPU core each. And the result was like this:
top - 14:37:54 up 2:35, 3 users, load average: 1.30, 3.60, 4.84
Tasks: 255 total, 1 running, 254 sleeping, 0 stopped, 0 zombie
%Cpu0 :100.0 us, 0.0 sy, 0.0 ni, 0.0 id, 0.0 wa, 0.0 hi, 0.0 si, 0.0 st
%Cpu1 : 1.7 us, 0.3 sy, 0.0 ni, 98.0 id, 0.0 wa, 0.0 hi, 0.0 si, 0.0 st
%Cpu2 : 41.6 us, 0.0 sy, 0.0 ni, 58.4 id, 0.0 wa, 0.0 hi, 0.0 si, 0.0 st
%Cpu3 :100.0 us, 0.0 sy, 0.0 ni, 0.0 id, 0.0 wa, 0.0 hi, 0.0 si, 0.0 st
%Cpu4 : 0.3 us, 0.0 sy, 0.0 ni, 99.7 id, 0.0 wa, 0.0 hi, 0.0 si, 0.0 st
%Cpu5 : 0.3 us, 0.3 sy, 0.0 ni, 99.3 id, 0.0 wa, 0.0 hi, 0.0 si, 0.0 st
%Cpu6 : 59.0 us, 0.0 sy, 0.0 ni, 41.0 id, 0.0 wa, 0.0 hi, 0.0 si, 0.0 st
%Cpu7 :100.0 us, 0.0 sy, 0.0 ni, 0.0 id, 0.0 wa, 0.0 hi, 0.0 si, 0.0 st
KiB Mem: 16369608 total, 11026436 used, 5343172 free, 62356 buffers
KiB Swap: 16713724 total, 360 used, 16713364 free. 2228576 cached Mem
PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND
10829 aaaaaa 20 0 42.624g 1.010g 142408 S 101.2 6.5 0:22.78 java
10861 aaaaaa 20 0 42.563g 1.044g 142340 S 101.2 6.7 0:22.75 java
10831 aaaaaa 20 0 42.704g 1.262g 142344 S 100.8 8.1 0:24.86 java
10857 aaaaaa 20 0 42.833g 1.315g 142456 S 100.5 8.4 0:26.48 java
1978 aaaaaa 20 0 1462096 186480 102652 S 1.0 1.1 0:34.82 compiz
10720 aaaaaa 20 0 7159748 1.579g 32008 S 1.0 10.1 0:16.62 java
1246 root 20 0 326624 101148 65244 S 0.7 0.6 0:50.37 Xorg
1720 aaaaaa 20 0 497916 28968 20624 S 0.3 0.2 0:02.83 unity-panel-ser
2238 aaaaaa 20 0 654868 30920 23052 S 0.3 0.2 0:06.31 gnome-terminal
I think java in the first 4 lines are Spark workers. If it's correct, it's nice that there are four Spark workers and each of them are using 1 physical core each (e.g., 101.2%).
But I see that 5 physical cores are used. Among them, CPU0, CPU3, CPU7 are fully used. I think one Spark worker is using one of those physical cores. It's fine.
However, the usage levels of CPU2 and CPU6 are 41.6% and 59.0%, respectively. They add up to 100.6%, and I think one worker's job is distributed to those 2 physical cores.
With SPARK_WORKER_INSTANCES=4 AND SPARK_WORKER_CORES=1, is this a normal situation? Or is this a sign of some errors or problems?
This is perfectly normal behavior. Whenever Spark uses term core it actually means either process or thread and neither one is bound to a single core or processor.
In any multitasking environment processes are not executed continuously. Instead, operating system is constantly switching between different processes which each one getting only small share of available processor time.
i'm running a CentOS 7.2 VM on Azure and get a CPU stuck kernel-bug warning. top shows that CPU#0 is 100% in use.
[admin#bench2 ~]$
Message from syslogd#bench2 at Feb 9 10:06:43 ...
kernel:BUG: soft lockup - CPU#0 stuck for 22s! [kworker/u128:1:13777]
This is the topoutput:
Tasks: 258 total, 7 running, 251 sleeping, 0 stopped, 0 zombie
%Cpu0 : 0.0 us,100.0 sy, 0.0 ni, 0.0 id, 0.0 wa, 0.0 hi, 0.0 si, 0.0 st
%Cpu1 : 0.3 us, 0.3 sy, 0.0 ni, 99.3 id, 0.0 wa, 0.0 hi, 0.0 si, 0.0 st
%Cpu2 : 0.0 us, 0.0 sy, 0.0 ni,100.0 id, 0.0 wa, 0.0 hi, 0.0 si, 0.0 st
%Cpu3 : 0.0 us, 0.0 sy, 0.0 ni,100.0 id, 0.0 wa, 0.0 hi, 0.0 si, 0.0 st
KiB Mem : 28813448 total, 26938144 free, 653860 used, 1221444 buff/cache
KiB Swap: 0 total, 0 free, 0 used. 27557900 avail Mem
PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND
73 root 20 0 0 0 0 S 0.7 0.0 1:03.03 rcu_sched
1 root 20 0 43668 6204 3796 S 0.0 0.0 0:04.70 systemd
2 root 20 0 0 0 0 S 0.0 0.0 0:00.03 kthreadd
3 root 20 0 0 0 0 R 0.0 0.0 0:00.10 ksoftirqd/0
5 root 0 -20 0 0 0 S 0.0 0.0 0:00.00 kworker/0:0H
Centos + Kernel Version:
CentOS Linux release 7.1.1503 (Core)
Linux bench2 3.10.0-229.7.2.el7.x86_64 #1 SMP Tue Jun 23 22:06:11 UTC 2015 x86_64 x86_64 x86_64 GNU/Linux
If noticed that this error also appears on CentOS 7.2 versions.
[84176.089080] BUG: soft lockup - CPU#0 stuck for 23s! [kworker/u128:1:13777]
[84176.089080] Modules linked in: vfat fat isofs xfs libcrc32c iptable_filter ip_tables udf crc_itu_t hyperv_fb hyperv_keyboard hv_utils i2c_piix4 i2c_core serio_raw pcspkr crct10dif_pclmul crc32_pclmul crc32c_intel ghash_clmulni_intel aesni_intel lrw gf128mul glue_helper ablk_helper cryptd ext4 mbcache jbd2 sd_mod crc_t10dif crct10dif_common hv_netvsc hv_storvsc hid_hyperv sr_mod cdrom ata_generic pata_acpi ata_piix libata floppy hv_vmbus
[84176.089080] CPU: 0 PID: 13777 Comm: kworker/u128:1 Tainted: G W -------------- 3.10.0-229.7.2.el7.x86_64 #1
[84176.089080] Hardware name: Microsoft Corporation Virtual Machine/Virtual Machine, BIOS 090006 05/23/2012
If this version does problems on Azure it is no problem to switch it. If this is the case, I want to know which CentOS version would be the best to run on an Azure environment.
I solved the problem by setting the Host caching on VHD to None. Odd behaviour but it works.
see screen here
I had the same issue, it's a disk performance issue (high IOPS/Latency etc.), not related to CPU or RAM (at least in my case).
The storage (NetApp) was very loaded, I solve it by moving to SSD, even using a large raid group with HDD (without a special load) didn't help.
We used a K8S setup, but I saw it on lot of CentOS with simple applications as well.
Regards,
I am doing network performance tests, and realized the interface's interrupts processing on 8 cpus is not balanced. So I want to make them more balanced.
I just set the files :
echo 11 > /proc/irq/16/smp_affinity
echo 22 > /proc/irq/17/smp_affinity
echo 44 > /proc/irq/18/smp_affinity
echo 88 > /proc/irq/19/smp_affinity
where the 16 17 18 and 19 are my four IRQ no of my network interfaces.
[root#localhost ~]# cat /proc/interrupts | grep ens
16: 30490 0 16838 427032 379 0 10678 0 IO-APIC-fasteoi vmwgfx, ens34, ens42
17: 799858 0 68176 0 78056 0 44715 0 IO-APIC-fasteoi ioc0, ens35, ens43, ens39
18: 2673 0 6149 0 7651 0 5585 0 IO-APIC-fasteoi uhci_hcd:usb2, snd_ens1371, ens40, ens44
19: 145769 1431206 0 0 0 0 305 0 IO-APIC-fasteoi ehci_hcd:usb1, ens41, ens45, ens33
But, sadly, I still found the IRQ is not balanced over the CPUs:
Tasks: 263 total, 2 running, 261 sleeping, 0 stopped, 0 zombie
%Cpu0 : 7.5 us, 10.0 sy, 0.0 ni, 65.3 id, 0.0 wa, 0.4 hi, 16.7 si, 0.0 st
%Cpu1 : 9.7 us, 15.0 sy, 0.0 ni, 59.1 id, 0.0 wa, 0.0 hi, 16.2 si, 0.0 st
%Cpu2 : 11.7 us, 21.6 sy, 0.0 ni, 66.7 id, 0.0 wa, 0.0 hi, 0.0 si, 0.0 st
%Cpu3 : 10.4 us, 16.6 sy, 0.0 ni, 66.0 id, 0.0 wa, 0.0 hi, 6.9 si, 0.0 st
%Cpu4 : 10.9 us, 24.5 sy, 0.0 ni, 64.5 id, 0.0 wa, 0.0 hi, 0.0 si, 0.0 st
%Cpu5 : 11.8 us, 29.4 sy, 0.0 ni, 58.8 id, 0.0 wa, 0.0 hi, 0.0 si, 0.0 st
%Cpu6 : 9.0 us, 19.8 sy, 0.0 ni, 71.3 id, 0.0 wa, 0.0 hi, 0.0 si, 0.0 st
%Cpu7 : 11.5 us, 22.6 sy, 0.0 ni, 65.9 id, 0.0 wa, 0.0 hi, 0.0 si, 0.0 st
So, why the IRQs not occur on all CPUs?
How can I balance the irq processing over all CPUs?
Closed. This question is off-topic. It is not currently accepting answers.
Want to improve this question? Update the question so it's on-topic for Stack Overflow.
Closed 10 years ago.
Improve this question
free -m
total used free shared buffers cached
Mem: 7974 6993 981 0 557 893
-/+ buffers/cache: 5542 2432
Swap: 2047 0 2047
You see that my system has used 5542MB memory, but when I use ps aux to check who uses it, I couldn't figure out.
ps aux | awk '$6 > 0{print $3, $4, $5, $6}'
%CPU %MEM VSZ RSS
0.0 0.0 10344 700
0.0 0.0 51172 2092
0.0 0.0 51172 1032
0.0 0.0 68296 1600
0.0 0.0 12692 872
0.0 0.0 33840 864
0.0 0.0 10728 376
0.0 0.0 8564 648
0.0 0.0 74856 1132
53.2 0.5 930408 45824
0.0 0.0 24236 1768
0.0 0.0 51172 2100
0.0 0.0 51172 1040
0.0 0.0 68296 1600
51.9 0.5 864348 42740
0.0 0.0 34360 2672
0.0 0.0 3784 528
0.0 0.0 3784 532
0.0 0.0 3784 528
0.0 0.0 3784 528
0.0 0.0 3784 532
0.0 0.0 65604 900
0.0 0.0 63916 832
0.0 0.0 94020 5980
0.0 0.0 3836 468
0.0 0.0 93736 4000
0.0 0.0 3788 484
0.0 0.0 3652 336
0.0 0.0 3652 336
0.0 0.0 3684 344
0.0 0.0 3664 324
0.0 0.0 19184 4880
0.0 0.0 3704 324
0.0 0.0 340176 1312
0.0 0.0 46544 816
0.0 0.0 10792 1092
0.0 0.0 3824 400
0.0 0.0 3640 292
0.0 0.0 3652 332
0.0 0.0 3652 332
0.0 0.0 3664 328
0.0 0.0 4264 1004
0.0 0.0 4584 2368
0.0 0.0 77724 3060
0.0 0.0 89280 2704
you see, that the sum of RSS is 152.484MB, the sum of VSZ is 3376.34MB, so I don't know who eat up the rest of the memory, the kernel?
From my system:
$ grep ^S[^wh] /proc/meminfo
Slab: 4707412 kB
SReclaimable: 4602900 kB
SUnreclaim: 104512 kB
These three metrics are data structures held by the slab alocator. While SUnreclaimable is, well, unreclaimable, SReclaimable is just like any other cache in the system - it will be made available to processes under memory pressure. Unfortunately free does not seem to take it into account, as mentioned in detail in this older answer of mine, and this part of memory can easily grow to several GB of memory...
If you really want to see how much memory your processes are using you could try going through the cache-emptying procedure described in my post - you can skip the swap-related parts, since your system does not appear to be using any swap memory anyway.