I think load average is not only measured by CPU workload. It can also increase when there is a disk I/O bottleneck. I usually encounter very high load averages during specific backup jobs of folders with many small files from mechanical drives while the actual CPU utilization is below 30%.
I liked your grocery store analogy. When I worked in customer support (defect support for a large commercial software product), we had spikes in activity (incoming customer calls) occasionally, but we were able to eventually catch up while the incoming volume was lower/more normal. Over time, the number of incoming calls increased (in part due to the, uh... evolving quality of the product and new releases, as well as increasing numbers of customers), and the "spikes" started to become the new "normal" volume. Unfortunately, management was very slow at hiring new people, so we got more and more overwhelmed, as our load average grew ever higher. Eventually people started leaving (CPU cores burning out? 😁 ), which made the load even higher on the remaining people. I've since moved on to other things, and I'm enjoying my work a lot more than I did when I was in that job. I do credit it for helping me to get into my later positions, though. Anyway, just wanted to say I did like your analogy, I thought it illustrates the concept well. I've been using Linux for 24 years now, and I knew a higher load average is "bad", or at least hints at a potentially overloaded system, but I always got confused about how to actually interpret it. I guess the idea that "1 means 100%" comes from the days of single-core systems. This video definitely helped me to understand what's really going on.
I've been doing some performance testing over the past week and you can get load averages that exceed the number of cpus. So loads above 100 percent by this measurement. I believe load avg is determined by the number of processes waiting to be completed so it shouldn't exceed number of cpus when working normally. But when ur slamming the system and there's serious bottlenecking that's not necessarily the case. The highest I saw was loads of around 3 on 2 cores (when I started seeing loads above 100 percent I got kinda interested and tried to see how high it could get). Ultimately load avg was more determined by the network interface on the machines than the cpus they had. Testing was done on pfsense firewalls and I was trying to see how low power consumption could be pushed without affecting performance. Firewalls are on all the time and spend a lot of that time at very low load if ur not servicing users at all hours.
Correct, the load average can go above the number of CPUs. I’ve seen it go to 500 and higher, to the point where the CPU is so pegged it can’t even authenticate an SSH session so you can even connect to it and fix it.
@@LearnLinuxTV I still think the 100 percent measurement is useful. As I understand it it's at that point that users will start noticing and making calls to IT
i have 2 cores and 2 threads per core so it makes 4 CPUs , so I have 4 CPUs , so does that mean i have 4 cores or 4 threads , and if it will have 4 threads then why are you saying you have 2 cores because youe have 1 core and 2 threads so that makes it 2 CPUs. So please explain this ?
Why, if you're running a 32-core EPYC CPU, did /proc/cpuinfo show 'cpu cores: 1" ? That makes no sense to me. For this VM, you have at least 2 cores. Shouldn't it say 2 or 32?
The percentage analogy is ALWAYS a bad one, when looking at load average. It is the average number of processes queued to be handled by cpu package, over a given length of time(1, 5 and 15 minutes as you mentioned). This is distinctly different from determining capacity, because it is NOT how many processes actually handled over that time. Only how many are queued for cpu, over that time. A load avg of 4 on a quad core cpu, doesn't mean the cpu is at 100% capacity. The way processes are managed in the kernel, means linux can adequately manage a number of processes well over the number of cpu cores. To know if the load is high, takes a much more holistic approach to troubleshooting, involving consideration of CPU states.
@Jay Just wanted to say you a great job with these vids Jay. And I also wanted to put a comment down just in case it truly does help with the algorithm stuff and your channel being promoted/recommended by GoogleTube..
I see @abu bua mentioned "nproc" as an alternate for processor info. "lscpu" can be used as well. It is usualy installed on most Linux distros, but can be installed as part of the "util-linux" package if not on your system. Thank you Jay... Awesome information. I have learned 80% of my Linux commands from watching your videos and reading comments.
Wait a min, you have 48 Cores for Desktop and calculations based on load average based with CPU... ok, Overkill. The load then, will never reach 48.0 unless it's trying to compile, maybe the Linux kernel 48 times at 1 go [maybe the only way to know is to try it] Peace :-)
Awesome as always Jay - Is it the number of CPUs or Cores that has to be reached in the load average for it to be 100% used. I noticed you said 2 for your example but your cpuinfo shows that you have a multicore system. Thanks!
