top of page
Recent Posts
Writer's pictureMunshi Hafizul Haque

Performance assessment and tuning experiences with RHEL

Performance management is the process of making sure that adequate computing resources (i.e. CPU, Memory, Disk and Networks) are available to accomplish the business needs of all users. Before performing any performance assessment we hope the system has design properly and have conducted below exercise:

  • Workload optimization: The target workload may be hampered by these settings.

  • Capacity Planning: To estimate the resources that will be necessary to support a system's workload for a specific period of time.

  • Throughput and Latency: Throughput is the measure of how much data can be transferred or processed by a resource in a given time. Latency is the delay that a resource must wait to start data transfer or processing as opposed to Throughput.

For the performance assessment and tuning, we have to follow some approaches. The Utilization Saturation and Errors (USE) Method that is highly regarded among performance tuning experts.


With Utilization Saturation and Errors (USE) Method, we can create our metric table. Below an example of very high level view of listing of resources, consider the metric types: utilization, saturation and errors and so on.

resource

type

metric

CPU

utilization

CPU utilization (either per-CPU or a system-wide average)

CPU

saturation

run-queue length or scheduler latency

Memory capacity

utilization

available free memory (system-wide)

Memory capacity

saturation

anonymous paging or thread swapping (maybe "page scanning" too)

Network interface

utilization

RX/TX throughput / max bandwidth

Storage device I/O

utilization

device busy percent

Storage device I/O

saturation

wait queue length

Storage device I/O

errors

device errors ("soft", "hard")

  • Utilization: 100% utilization is usually a sign of a bottleneck (check saturation and its effect to confirm the bottleneck). Greater than 70% utilization for an extended time (many seconds or minutes) can hide short bursts of 100% utilization.

  • Saturation: Any degree of saturation can be a problem (non-zero) and is measured as the length of a wait queue, or the time spent waiting in the queue before being processed.

  • Errors: Non-zero error counters are worth investigating, especially if they are still increasing while facing degradation in performance.

Note: More details refere to: https://queue.acm.org/detail.cfm?id=2413037

Profiling:

Profiling is the process of gathering information in different ways of retrieving performance data for a system. For example, Application Profiling is gathering information about a program's behavior as it executes. And will determine which areas of a program can be optimized to increase the program's overall speed, reduce its memory usage, etc. Application profiling tools help to simplify this process.


In a nutshell, we need to prepare the set of tools to gathering information performance data for a system.


System monitoring:

There are a bunch of monitoring tools that commonly view information, and can be used by way of the command line or a graphical user interface, as determined by the system administrator. And system monitoring is a helpful approach to provide the user with data regarding the actual timing behavior of the system to perform further analysis using the data that these monitors provide.

It’s always advisable to use more than one monitoring tool or alternative tool to verify the data regarding the actual timing behavior of the system.


Process Management in Linux:

Whatever program that we execute in our linux system will consider process and we need to understand the type of process (Foreground or Background), states of the process (Running, Sleeping, etc.), resource utilization (CPU,Memory, etc) and so on.


All possess information about will be found under /proc directory and proc - process information pseudo-filesystem. More details on /proc refer to the man page of proc, or “man proc” commands.


Linux processes types:

  • Foreground Processes: a interactive processes and depend on the user for input

  • Background Processes: A non-interactive or automatic process and runs independently of the user for any input.

Process States in Linux:

'R' = RUNNING & RUNNABLE

  • A runnable process is ready and lined up to be run, but for whatever reason, the CPU is not ready for it to be scheduled.

  • A running program is actively running and allocated to a CPU/CPU core or thread.

'D' = UNINTERRUPTABLE_SLEEP

  • UNINTERRUPTABLE_SLEEP is a state where the process is waiting on something as well. But typically in this state, interrupting could cause some major issues. It is rare to catch a process in this state but when it is, it is usually due to a system call or syscall. Typically a process in UNINTERRUPTABLE_SLEEP will not wake-up.

'S' = INTERRRUPTABLE_SLEEP

  • During the course of a process running, it will get to a point where it is waiting on data. This may be in the form of input from the terminal such as asking the user for input. And this process in INTERRRUPTABLE_SLEEP will wake-up to handle signals.

'T' = STOPPED

  • The STOPPED process you might think of it more as a suspended process and A process enters a stopped state when it receives a stop signal or uses Control + Z.

'Z' = ZOMBIE

  • Processes in ZOMBIE state may sound like a strange state to be in. In basic terms, this is an interim state after a process exits — but before its parent removes it from the process table. That is, Zombie state is when a process is dead but the entry for the process is still present in the table.

What we are looking into is the running process in the Linux system. For example:

  • Load Average

  • CPU Usages at system and user space level

  • CPU or Memory Intensive Process

  • Number of processes that in ZOMBIE state

  • Number of processes that in UNINTERRUPTABLE_SLEEP state

  • High number of context switching

  • High number of context switching

  • Number of runnable processes (running or waiting for run time).

  • Number of processes blocked waiting for I/O to complete.

Let's use some commands that are available in Linux to track running processes.

  • pidstat - Report statistics for Linux tasks.

  • top - display Linux processes

  • ps - report a snapshot of the current processes.

  • pstree - display a tree of processes

  • pstack - print a stack trace of a running process

  • pmap - report memory map of a process

  • sar - Collect, report, or save system activity information

  • vmstat - Report virtual memory statistics

  • mpstat - Report processors related statistics.

Note: sysstat, psmisc, procps-ng, memstrack and gdb rpm has to be installed in your system.

Below command shows load average based on 1 minute, 5 minutes and 15 minutes.

# cat /proc/loadavg
0.33 0.37 0.53 2/2058 613675
# w
 22:42:15 up 1 day,  1:30,  1 user,  load average: 0.30, 0.36, 0.53
USER 	TTY  	FROM         	LOGIN@   IDLE   JCPU   PCPU WHAT
mhaque   :0   :0           Mon21   ?xdm?   2:45m  0.00s /usr/libexec/gdm-x-session --register-session --run-script gnome-sessio
# sar -q -f /var/log/sa/sa13
09:20:25 PM   runq-sz  plist-sz   ldavg-1   ldavg-5  ldavg-15   blocked
09:30:17 PM     	0  	1338  	0.73  	0.71  	0.51     	1
09:40:14 PM     	0  	1342  	0.55  	0.56  	0.50     	0
09:50:25 PM     	0  	1318  	0.32  	0.41  	0.45     	0
10:00:25 PM     	0  	1322  	0.10  	0.15  	0.28     	0
10:10:14 PM     	0  	1302  	0.37  	0.39  	0.34     	0
10:20:25 PM     	0  	1357  	0.37  	0.39  	0.36     	0
10:30:25 PM     	0  	1349  	0.55  	0.53  	0.45     	0
10:40:14 PM     	0  	1755  	0.27  	0.79  	0.70     	0
10:50:25 PM     	0  	1740  	0.85  	0.59  	0.61     	0
11:00:25 PM     	0  	1754  	0.01  	0.22  	0.44     	0
11:10:14 PM     	0  	1754  	0.05  	0.09  	0.25     	0
11:20:25 PM     	1  	1747  	0.09  	0.08  	0.16     	0
11:30:25 PM     	0  	1729  	0.02  	0.08  	0.14     	0
11:40:14 PM     	0  	1738  	0.17  	0.08  	0.09     	0
11:50:25 PM     	1  	1727  	0.35  	0.24  	0.14     	0
Average:        	0  	1551  	0.32  	0.35  	0.36     	0
# top

Below command shows CPU Utilization (user, system,iowait,idle), context switching etc. and we need to check the RED columns for verify utilization.

# sar -u -f /var/log/sa/sa13
# sar -P 2 -f /var/log/sa/sa13
# sar -P ALL -f /var/log/sa/sa13
Average:    	CPU 	%user 	%nice   %system   %iowait	%steal 	%idle
Average:    	all  	1.06  	0.01  	 0.39  	0.01  	0.00 	    98.53
Average:      0  	0.97  	0.00  	 0.37  	0.01  	0.00 	    98.65
Average:      1  	1.44  	0.00  	 0.33  	0.01  	0.00 	    98.22
Average:      2  	1.10  	0.00  	 0.28  	0.01  	0.00 	    98.61
Average:      3  	1.16  	0.02  	 0.30  	0.01  	0.00 	    98.50
Average:      4  	1.21  	0.00  	 1.18  	0.02  	0.00 	    97.60
Average:      5  	1.00  	0.00  	 0.29  	0.02  	0.00 	    98.69
Average:      6  	0.99  	0.01  	 0.28  	0.01  	0.00 	    98.71
# vmstat 2 5
procs -----------memory---------- ---swap-- -----io---- -system-- ------cpu-----
 r  b   swpd   free   buff  cache   si   so	bi	bo   in   cs us sy id wa st
 1  0   0 89754592   4048 22219692	0	0 	7	21	1	2  1  0 99  0  0
 1  0   0 89764976   4048 22205664	0	0 	0	26 4308 6268  1  1 98  0  0
 0  0   0 89781160   4048 22188544	0	0 	0 	4 3516 5708  1  0 98  0  0
 1  0   0 89779152   4048 22188608	0	0 	0	88 3777 5750  1  1 98  0  0
 1  0   0 89776576   4048 22165924	0	0 	0   154 4569 6985  2  1 98  0  0
# sar -w -f /var/log/sa/sa13
09:20:25 PM	    proc/s   cswch/s
09:30:17 PM  	5.27   7501.50
09:40:14 PM  	4.21   6767.55
09:50:25 PM  	4.28   5564.40
10:00:25 PM  	3.92   4990.68
10:10:14 PM  	4.97   6218.96
10:20:25 PM  	6.08   5890.19
10:30:25 PM  	5.49   7829.43
# sar -q -f /var/log/sa/sa12
07:20:08 PM   runq-sz  plist-sz   ldavg-1   ldavg-5  ldavg-15   blocked
07:30:09 PM     	0  	2490  	0.93  	0.78  	0.70     	0
07:40:09 PM     	0  	2495  	0.63  	0.47  	0.56     	0
07:50:09 PM     	0  	2500  	0.34  	0.29  	0.40     	0
08:00:09 PM     	1  	2496  	0.38  	0.25  	0.29     	0
08:10:09 PM     	1  	2514  	0.87  	0.92  	0.65     	0
08:20:09 PM     	0  	2508  	0.07  	0.28  	0.46     	0
10:20:09 PM     	0  	2378  	0.19  	0.15  	0.18     	0
Average:        	0  	2319  	0.23  	0.23  	0.22     	0
# mpstat  2 10
Linux 4.18.0-372.19.1.el8_6.x86_64 (munshi-lab.jazakallah.info)     02/15/2023     _x86_64_    (16 CPU)
09:17:18 AM  CPU	%usr   %nice	%sys %iowait	%irq   %soft  %steal  %guest  %gnice   %idle
09:17:20 AM  all	0.53	0.00	0.31	0.00	0.13	0.03	0.00	0.16	0.00   98.84
09:17:22 AM  all	1.12	0.00	0.37	0.00	0.16	0.12	0.00	0.22	0.00   98.00
09:17:24 AM  all	0.88	0.00	0.31	0.03	0.06	0.06	0.00	0.19	0.00   98.47
09:17:26 AM  all	1.09	0.00	0.47	0.03	0.12	0.00	0.00	0.53	0.00   97.75
09:17:28 AM  all	1.00	0.00	0.37	0.00	0.12	0.12	0.00	0.22	0.00   98.16
09:17:30 AM  all	0.94	0.00	0.31	0.00	0.16	0.03	0.00	0.16	0.00   98.41

Below command can show you the current context switching statistics per process in case of of any high number of context switching occur. And we need to check the RED columns for verify utilization.

#  grep ctxt /proc/$PID/status
Note: $PID is the possess ID and it must be a number.

# grep ctxt /proc/9124/status
voluntary_ctxt_switches:	136922
nonvoluntary_ctxt_switches:	210
# pidstat -wt 3 10
10:59:32 PM   UID  	TGID   	TID   cswch/s nvcswch/s  Command
10:59:35 PM 	0     	1     	-  	0.33  	0.00  systemd
10:59:35 PM 	0     	-     	1  	0.33  	0.00  |__systemd
10:59:35 PM 	0    	12     	-  	2.62  	0.00  ksoftirqd/0
10:59:35 PM 	0     	-    	12  	2.62  	0.00  |__ksoftirqd/0
10:59:35 PM 	0    	13     	-	120.66  	0.00  rcu_sched
Note: Why does the system show a high number of context switching and interrupt rate?: https://access.redhat.com/solutions/69271

Memory Management in Linux:

Memory Management is one of the more sophisticated things that the kernel does.The

Computer systems organize memory into fixed-size chunks called pages. The default size of a page is 4 KiB on the x86_64 processor architecture.

The size of a process's virtual address space does not depend on the installed physical RAM but rather depends on the processor architecture. On a 64-bit x86-64 system, the address space is 2 64 bytes (16 EiB) in size.


Tools such as ps and top distinguish between two statistics:

  • VIRT (or VSZ) — the total amount of virtual memory a process has asked for,

  • RES (or RSS) — the total amount of virtual memory that a process is currently mapping to physical memory.

The Virtual Memory is a memory management technique that is implemented using both hardware (MMU) and software (operating system) and we have to understand a few terminology during the course of a virtual memory allocation.

  • TLB lookup: A virtual address needs to be translated into a physical address, the MMU first searches for it in the TLB cache.

  • TLB Hit: The physical address is returned and the computation simply goes on.

  • TLB Miss: No match for the virtual address in the TLB cache and the MMU searches for a match on the whole page table.

  • TLB Update: If this match exists on the page table, the address translation is restarted so that the MMU is able to find a match on the update.

The page supervisor is a software component of the operating system kernel that typically raises a segmentation fault exception or a page fault occurs. which means the requested page has to be retrieved from the secondary storage (i.e., disk) where it is currently stored. And to move the page from main memory to disk, the paging supervisor may use several page replacement algorithms, such as Least Recently Used (LRU).

  • Page Fault: A page table lookup may fail due to either no valid translation for the specified virtual address or the requested page is not loaded in main memory at the moment. And the page has to be retrieved from the secondary storage (i.e., disk)

  • Page Swap: To move the page from main memory to disk, the paging supervisor may use several page replacement algorithms, such as Least Recently Used (LRU).

    • swapping-out: The process of writing pages out to disk to free memory is called swapping-out.

    • swapping-in: The kernel will read back in the page from the disk and satisfy the page fault is called swapping-in.

What we are looking into is the memory usilization for running process in the Linux system. For example:

  • Average Memory utilization

  • CPU or Memory Intensive Process

  • Swapping-in and swapping-out ratio

  • To review the amount of active, inactive, and dirty memory, inspect the /proc/meminfo file.

    • Free — The page is available for immediate allocation.

    • Active — The page is in active use and not a candidate for being freed.

    • Inactive clean — The page is not in active use, and its content matches the content on disk.

    • Inactive dirty — The page is not in active use, but the page content has been modified since being read from disk and has not yet been written back.

  • To review memory overcommit policies

  • Number of minor page faults and major page faults

Let's use some commands that are available in Linux to track running processes.

  • top - display Linux processes

  • ps - report a snapshot of the current processes.

  • sar - Collect, report, or save system activity information

  • vmstat - Report virtual memory statistics

  • free - Display amount of free and used memory in the system

  • memstrack - To analyze the memory usage of a certain program/module/code.

  • hwloc-gui - lstopo, lstopo-no-graphics, hwloc-ls - Show the topology of the system

  • hwloc - lstopo, lstopo-no-graphics, hwloc-ls - Show the topology of the system

Note: in addition memstrack, hwloc-gui, and hwloc rpm have to be installed in your system.

The command below will help us to get memory utilization, like Free, Avaiable, Buffer, Cache etc. And alos will show Swap In & Out statistics. We need to check the RED columns for verify utilization.

# free -m
          	total    	used    	free  	shared  buff/cache   available
Mem:     	128168   	19642   	86670    	1767   	21855  	105538
Swap:      	8191       	0    	8191
Note: What is the difference between cache and buffer ?https://access.redhat.com/solutions/636263
# vmstat 2 3
procs -----------memory---------- ---swap-- -----io---- -system-- ------cpu-----
 r  b   swpd   free   buff  cache   si   so	bi	bo   in   cs us sy id wa st
 2  0  	0 88751552   4048 22377660	0	0 	7	21   11   17  1  0 98  0  0
 2  0  	0 88751552   4048 22377660	0	0 	7	21   11   17  1  0 98  0  0
 2  0  	0 88751552   4048 22377660	0	0 	7	21   11   17  1  0 98  0  0
# cat /proc/meminfo
MemTotal:   	131245016 kB
MemFree:    	88770688 kB
MemAvailable:   108090688 kB
Buffers:        	4048 kB
Cached:     	20861156 kB
SwapCached:        	0 kB
Active:      	9339204 kB
Inactive:   	28706308 kB
Active(anon):	2749484 kB
Inactive(anon): 16248396 kB
Active(file):	6589720 kB
Inactive(file): 12457912 kB
Unevictable: 	1076304 kB
Mlocked:          	16 kB
SwapTotal:   	8388604 kB
SwapFree:    	8388604 kB
Dirty:           	796 kB
Writeback:         	0 kB
# sar -r -f /var/log/sa/sa12
07:20:08 PM kbmemfree   kbavail kbmemused  %memused kbbuffers  kbcached  kbcommit   %commit  kbactive   kbinact   kbdirty
07:30:09 PM  97075468 114306356  34169548 	26.03  	3840  20045848  43739880 	31.32   5278384  24181652   	720
07:40:09 PM  97099800 114329872  34145216 	26.02  	3840  20028924  43935408 	31.46   5278060  24190224  	1348
07:50:09 PM  96912188 114153088  34332828 	26.16  	3840  20065144  44059880 	31.55   5280116  24348516  	1848
08:00:09 PM  96192160 113434132  35052856 	26.71  	3840  20644384  44663516 	31.99   5282572  25060108  	4880
08:10:09 PM  96936512 114640448  34308504 	26.14  	3840  20450336  43657280 	31.27   5282272  24340980  	3828
08:20:09 PM  96191264 114519864  35053752 	26.71  	3840  21005072  43609060 	31.23   5280768  25227244  	2196
08:30:09 PM  96334216 114677760  34910800 	26.60  	3840  20941680  43564988 	31.20   5282856  25131884   	128
Average: 	99581906 116030537  31663110 	24.13  	3840  18912848  38508292 	27.58   5080471  22259235  	1701
# sar -W -f /var/log/sa/sa12
07:20:08 PM  pswpin/s pswpout/s
07:30:09 PM  	0.00  	0.00
07:40:09 PM  	0.00  	0.00
07:50:09 PM  	0.00  	0.00
08:00:09 PM  	0.00  	0.00
08:10:09 PM  	0.00  	0.00
08:20:09 PM  	0.00  	0.00
08:30:09 PM  	0.00  	0.00
08:40:09 PM  	0.00  	0.00
08:50:09 PM  	0.00  	0.00
Average:     	0.00  	0.00

There are some useful command to identify utilization and status per process :

To check CPU and memory incentive per process

# ps axo %cpu,%mem,pid,user,args --sort %cpu
# ps axo %cpu,%mem,pid,user,args --sort %mem

To check total virtual memory requested and total physical memory mapped per process:

# ps axo pid,rsz,vsz,user,args --sort vsz
# ps axo pid,rsz,vsz,user,args --sort rsz

To view minor and major page faults per process:

# ps axo  pid,minflt,majflt,user,args --sort majflt
# ps axo  pid,minflt,majflt,user,args --sort minflt

To check processes with D state in your system:

# ps auxH | awk '$8 ~ /^D/{print}'
root  	568098  0.0  0.0  	0 	0 ?    	D<   21:25   0:00 [kworker/u33:1+i915_flip]

To check processes with Z state in your system:

# ps auxH | awk '$8 ~ /^Z/{print}'
mhaque  	9246  0.0  0.0  	0 	0 tty2 	Z+   Feb13   0:00 [sd_cicero] <defunct>
Note: To check the PROCESS STATE CODES & STANDARD FORMAT SPECIFIERS section in “man ps” for more details on process status.

Note: some guide line to check system while unreachable/unresponsible: https://access.redhat.com/solutions/661503

Disk I/O Management in Linux:

Most legacy general tuning procedures for hard disks do not apply to SSDs. For example, SSDs do not require the use of read ahead and write behind caches. Caches should be configured as write-through. Red Hat does not recommend using journaling on SSD devices, because of increased SSD wear and the slowness caused by the unnecessary double writing.


What we are looking into is the running process in the Linux system. For example:

  • Average Memory utilization

  • CPU or Memory Intensive Process

  • Swapping-in and swapping-out ratio

  • To review the amount of active, inactive, and dirty memory, inspect the /proc/meminfo file.

    • Free — The page is available for immediate allocation.

    • Active — The page is in active use and not a candidate for being freed.

    • Inactive clean — The page is not in active use, and its content matches the content on disk.

    • Inactive dirty — The page is not in active use, but the page content has been modified since being read from disk and has not yet been written back.

  • To review memory overcommit policies

  • Number of minor page faults and major page faults

Let's use some commands that are available in Linux to track running processes.

  • top - display Linux processes

  • iostat - report a snapshot of the current processes.

  • sar - Collect, report, or save system activity information

  • iotop - simple top-like I/O monitor

Note: In addition iotop rpm have to be installed in your system.

What we are looking into is the running process in the Linux system. For example:

  • Average Disk utilization

  • To verify the low CPU usage (us field), and high CPU wait (wa field) scenario on top command.

# iostat -x 2 3
Linux 4.18.0-372.19.1.el8_6.x86_64 (munshi-lab.jazakallah.info) 	02/15/2023 	_x86_64_	(16 CPU)

avg-cpu:  %user   %nice %system %iowait  %steal   %idle
           1.20    0.00    0.40    0.02    0.00   98.38

Device            r/s     w/s     rkB/s     wkB/s   rrqm/s   wrqm/s  %rrqm  %wrqm r_await w_await aqu-sz rareq-sz wareq-sz  svctm  %util
nvme0n1          1.30   15.09     67.02    298.76     0.01     0.76   0.47   4.78    0.35    1.58   0.02    51.44    19.80   0.46   0.76
dm-0             1.13   15.26     66.61    298.74     0.00     0.00   0.00   0.00    0.30    4.17   0.06    58.96    19.57   0.47   0.76
dm-1             0.39    1.45     33.21     23.40     0.00     0.00   0.00   0.00    0.32    2.73   0.00    84.24    16.10   0.66   0.12
dm-2             0.00    0.00      0.01      0.00     0.00     0.00   0.00   0.00    0.19    0.00   0.00    22.65     0.00   0.15   0.00
dm-3             0.56   12.53     12.83    243.80     0.00     0.00   0.00   0.00    0.25    4.54   0.06    23.00    19.46   0.42   0.55
dm-4             0.18    1.07     20.55     31.54     0.00     0.00   0.00   0.00    0.42    2.36   0.00   116.07    29.45   1.02   0.13
Note: we can check any spesific disk utilization by "iostat -xd <disk name>. For Example: # iostat -xd sda
# iotop

# sar -d -f /var/log/sa/sa12
09:40:09 PM   	DEV   	tps 	rkB/s 	wkB/s   areq-sz	aqu-sz 	await 	svctm 	%util
09:50:09 PM  dev259-0  	8.26  	0.00	137.69 	16.67  	0.01  	1.41  	0.57  	0.47
09:50:09 PM  dev253-0  	8.22  	0.00	137.69 	16.75  	0.03  	3.05  	0.58  	0.48
09:50:09 PM  dev253-1  	0.79  	0.00  	4.37  	5.53  	0.00  	1.73  	1.04  	0.08
09:50:09 PM  dev253-2  	0.00  	0.00  	0.00  	0.00  	0.00  	0.00  	0.00  	0.00
09:50:09 PM  dev253-3  	7.27  	0.00	133.32 	18.33  	0.02  	3.22  	0.56  	0.40
09:50:09 PM  dev253-4  	0.00  	0.00  	0.00  	0.00  	0.00  	0.00  	0.00  	0.00
Average: 	dev253-1  	0.91  	0.24  	8.50  	9.65  	0.00  	1.99  	0.95  	0.09
Average: 	dev253-2  	0.00  	0.00  	0.00  	0.00  	0.00  	0.00  	0.00  	0.00
Average: 	dev253-3 	12.70  	0.35	214.78 	16.94  	0.04  	3.15  	0.41  	0.52
Average: 	dev253-4  	0.00  	0.00  	0.00  	0.00  	0.00  	0.00  	0.00  	0.00

Network Management in Linux:

There is no generic configuration that can be broadly applied to every system for network performance. Below articles may help you to understand more on Red Hat Enterprise Linux Network Performance Tuning.


Red Hat Enterprise Linux Network Performance Tuning Guide

How to tune `net.core.netdev_max_backlog` and `net.core.netdev_budget` sysctl kernel tunables?


Linux kernel automatically adjusts the size of these buffers based on the current network utilization, but within the limits specified by kernel tunables that are related to the networking buffers (or queues) consist of the core networking read and write buffers (used for UDP and TCP), per-socket TCP read and write buffers, fragmentation buffers, and DMA buffers for the network card. And we can visualized by below symbolic page diagram.


What we are initially looking into is the network performance in the Linux system. For example:

  • The adapter firmware level

  • The calculation called the bandwidth delay product (BDP)

  • To Identifying the network bottleneck

    • Observe drops in ethtool -S ethX statistics

    • The Linux kernel, IRQs or SoftIRQs by Check /proc/interrupts and /proc/net/softnet_stat

    • The protocol layers IP, TCP, or UDP and uses netstat -s and looks for error counters.


How to check firmware version of NICs:

# ethtool -i eth0
driver: e1000e
version: 4.18.0-372.19.1.el8_6.x86_64
firmware-version: 0.4-4
expansion-rom-version: 
bus-info: 0000:00:1f.6
supports-statistics: yes
supports-test: yes
supports-eeprom-access: yes
supports-register-dump: yes
supports-priv-flags: yes
$ modinfo e1000e|head -5
filename:       /lib/modules/4.18.0-372.19.1.el8_6.x86_64/kernel/drivers/net/ethernet/intel/e1000e/e1000e.ko.xz
version:        4.18.0-372.19.1.el8_6.x86_64
license:        GPL v2
description:    Intel(R) PRO/1000 Network Driver
author:         Intel Corporation, <linux.nics@intel.com>
One of the recent example for buggy NIC driver causing an issues and it's slow POD to POD throughput via vxlan when the Hypervisor has bnx2x_en NIC(s): https://access.redhat.com/solutions/5921451

How to calculate the bandwidth delay product (BDP) that is used to verify the buffers are correctly sized. The ping command can be used to find the average round trip time.

# ping 192.168.121.1
PING 192.168.121.1 (192.168.121.1) 56(84) bytes of data.
64 bytes from 192.168.121.1: icmp_seq=1 ttl=64 time=0.029 ms
64 bytes from 192.168.121.1: icmp_seq=2 ttl=64 time=0.033 ms
64 bytes from 192.168.121.1: icmp_seq=3 ttl=64 time=0.033 ms
64 bytes from 192.168.121.1: icmp_seq=4 ttl=64 time=0.035 ms
64 bytes from 192.168.121.1: icmp_seq=5 ttl=64 time=0.043 ms
64 bytes from 192.168.121.1: icmp_seq=6 ttl=64 time=0.030 ms
64 bytes from 192.168.121.1: icmp_seq=7 ttl=64 time=0.028 ms
^C
--- 192.168.121.1 ping statistics ---
7 packets transmitted, 7 received, 0% packet loss, time 6128ms
rtt min/avg/max/mdev = 0.028/0.033/0.043/0.004 ms

In this example, an average round trip time (rtt) of 0.033 ms, or 0.000033 seconds. With a network speed (capacity) of 1 Gigabit per second:

1 Gb/s × 19.2 ms =
10000000000 b/s × 0.000033 s = 330000 b
341000 bits × 1/8 B/b = 41250 B
41250 Bytes = 40.28 KiB

This example results in a bandwidth delay product (BDP) of 40.28 KiB. And we need to rememeber that:

  • If the BDP goes above 64 KiB, TCP connections can utilize window scaling.

  • A TCP window is the amount of data sent to the remote system that has not yet been acknowledged.

  • If unacknowledged data grows to the window size, the sender will stop sending until previous data has been acknowledged.

To Identifying the network bottleneck:

$ ethtool eth0 
Settings for eth0:
	Supported ports: [ TP ]
	Supported link modes:   10baseT/Half 10baseT/Full
	                        100baseT/Half 100baseT/Full
	                        1000baseT/Full
	Supported pause frame use: No
	Supports auto-negotiation: Yes
	Supported FEC modes: Not reported
	Advertised link modes:  10baseT/Half 10baseT/Full
	                        100baseT/Half 100baseT/Full
	                        1000baseT/Full
	Advertised pause frame use: No
	Advertised auto-negotiation: Yes
	Advertised FEC modes: Not reported
	Speed: 1000Mb/s
	Duplex: Full
	Auto-negotiation: on
	Port: Twisted Pair
	PHYAD: 1
	Transceiver: internal
	MDI-X: on (auto)
netlink error: Operation not permitted
        Current message level: 0x00000007 (7)
                               drv probe link
	Link detected: yes
# ethtool -S eth0|grep -i error
 	rx_errors: 0
 	tx_errors: 0
 	rx_length_errors: 0
 	rx_over_errors: 0
 	rx_crc_errors: 0
 	rx_frame_errors: 0
 	rx_missed_errors: 0
 	tx_aborted_errors: 0
 	tx_carrier_errors: 0
 	tx_fifo_errors: 0
 	tx_heartbeat_errors: 0
 	tx_window_errors: 0
 	rx_long_length_errors: 0
 	rx_short_length_errors: 0
 	rx_align_errors: 0
 	rx_csum_offload_errors: 13
 	uncorr_ecc_errors: 0
 	corr_ecc_errors: 0
# netstat -s|grep err
	0 packet receive errors
	0 receive buffer errors
	0 send buffer errors
$ netstat -i
Kernel Interface table
Iface             MTU    RX-OK RX-ERR RX-DRP RX-OVR    TX-OK TX-ERR TX-DRP TX-OVR Flg
eth0             1500  4050869      0     16 0       3320532      0      0      0 BMRU
lo              65536    34016      0      0 0         34016      0      0      0 LRU
virbr0           1500   334390      0      0 0        498501      0      0      0 BMRU
virbr1           1500        0      0      0 0             0      0      0      0 BMU
vnet3            1500  1222835      0      0 0       1136447      0      0      0 BMRU
vnet4            1500   789056      0      0 0       1211325      0      0      0 BMRU
wlan0            1500   290440      0      0 0        103798      0      0      0 BMRU
#  egrep "CPU0|eth0" /proc/interrupts
        	CPU0   	CPU1   	CPU2   	CPU3   	CPU4   	CPU5   	CPU6   	CPU7   	CPU8   	CPU9   	CPU10  	CPU11  	CPU12  	CPU13  	CPU14  	CPU15 	 
 151:      	0      	0    	170   	2531      	0    	702    	830      	0     	13      	0	5548270  	62825   	1978  	67774   	2584    	184  IR-PCI-MSI 520192-edge  	eth0

Look into all above the error and drop packets numbers.


Note: We can also use below command to verify network performance.
# sar -n DEV 1
# sar -n TCP,ETCP 1
Note: Some usefull reference for Network performance debugging: How can I tune the TCP Socket Buffers: https://access.redhat.com/solutions/369563 & How to begin Network performance debugging: https://access.redhat.com/articles/1311173

Conclution, all the above tools will give a bunch of data that need to be analyzed. But the collecting date could be historical data or point in time data.


Point in time data is good for analyzing issues that are ongoing in the system.

historical data will be collected periodically in the background and can analyze issues that happened based on timestamp assumption.


We can configure sar (by default in REL8), or nmon and any other tools in the background to collect performance data and store in a location. For example: sar command will store data in binary format under the “/var/log/sa”.


Next action plan will be the tuning after analyzing performance data and the suspected bottleneck.

369 views0 comments

Recent Posts

See All
Log In to Connect With Members
View and follow other members, leave comments & more.

JazakAllah© 2020. All Rights Reserved

keep growing your community bigger, Keep your IT knowledge up-to-date.

  • LinkedIn
bottom of page