It should also be mentioned, Linux Load Average is a complex beast[1]. However, a general rule of thumb that works for most environments is:
You always want the load average to be less than the total number of CPU cores. If higher, you're likely experiencing a lot of waits and context switching.
On Linux this is not true, on an IO heavy system - with lots of synchronous I/Os done concurrently by many threads - your load average may be well over the number of CPUs, without having a CPU shortage. Say, you have 16 CPUs, load avg is 20, but only 10 threads out of 20 are in Runnable (R) mode on average, and the other 10 are in Uninterruptible sleep (D) mode. You don't have a CPU shortage in this case.
Note that synchronous I/O completion checks for previously submitted asynchronous I/Os (both with libaio and io_uring) do not contribute to system load as they sleep in the interruptible sleep (S) mode.
That's why I tend to break down the system load (demand) by the sleep type, system call and wchan/kernel stack location when possible. I've written about the techniques and one extreme scenario ("system load in thousands, little CPU usage") here:
The proper way is to have a idea of what it normally is before you need to troubleshoot issues.
What is a 'good load' depends on the application and how it works. Some servers something close to 0 is a good thing. Other servers a 10 or lower means something is seriously wrong.
Of course if you don't know what is a 'good' number or you are trying to optimize a application and looking for bottlenecks then it is time to reach for different tools.
If you like this post, I would recommend “BPF Performance Tools” and “Systems Performance: Enterprise and the Cloud” by Brenden Gregg.
I have pulled out a few miracles using these tools (identifying kernel bottlenecks or profiling programs using ebpf) and it has been well worth the investment to read through the books.
Literally did miracles at my last job with the first book and that got me my current job, where I also did some impressive proving which libraries had what performance with it again... Seriously valuable stuff.
Yea it is kindof cheating. I was helping someone debug why their workload was soft locking. I ran the profiling tools and found that cgroup accounting for the workload was taking nearly all the cpu time on locks. From searches through linux git logs I found that cgroup accounting in older kernels had global locks. I saw that newer kernels didn’t have this, so we moved to a newer kernels and all the issues went away.
Kernel/distro upgrades can cause severe regressions. Generally the approach we do is have the service run a canary which runs on the newer kernel for a while to A/B test the upgrade. Generally we rely on the service owner to validate this A/B test as we don’t want to own making sure services are healthy on the new kernel. This means it is primarily on the service owner to look at the results of the A/B test to determine if the upgrade is ok.
I build a cloud (such as AWS) and we have many tenants running on the cloud. Much like AWS will not force upgrade AMIs, we will not force tenants to upgrade either.
It is excellent and contains most things you could need. Downside is that it isn't yet a standard tool so you need to get it installed across your fleet
The iostat command has always been important to observe HDD/SDD latency numbers.
Especially SSDs are treated like magic storage devices with infinite IOPS at Planck-scale latency.
Until you discover that SSDs that can do 10GB/s don't do nearly so well (not even close) when you access them in a single thread with random IOPS, with queue depth of 1.
That's where you start down the eBPF rabbit hole with bcc/biolatency and other block device histogram tools. Further, the cache hit rate and block size behavior of the SSD/NVME drive can really affect things if, say, your autonomous vehicle logging service uses MCAP with a chunk size much smaller than a drive block... Ask me how I know
After this article was written, `free -m` on many systems started to have an "available" column that shows the sum of reclaimable and free memory. It's nicer than the "-/+" section shown in this old article.
$ free -m
total used free shared buff/cache available
Mem: 3915 2116 1288 41 769 1799
Swap: 974 0 974
Why did they not choose to use it for both (or neither)? I.e., what reasons for using FreeBSD on CDN servers would not also apply to using them for API servers?
They are extremely different workloads so.. everything?
The CDN servers are basically appliances, and are often embedded in various data centers (includes those ran by ISP's) to aggressively cache content. They care about high throughput and run a single workload. Being able to fine tune the entire stack, right down to the TCP/IP implementation is very valuable in this case. Since they ship the hardware and software, they can tightly integrate the two.
By contrast, API workloads are very heterogeneous. I'd have to imagine the ability to run any standard Linux software there would also be a big plus. Linux clearly has much more vetting on cloud providers than FreeBSD as well.
Can't you fine tune linux as well? Does FreeBSD perform better somehow on a CDN workload? I find it difficult to imagine that the reason is performance. But I don't know what the reason is.
tl;dw: the performance, the efficiency of development, the community, FreeBSD is a complete operating system, the code base is smaller, the ports system, and the license.
Also potentially a reason: According to drewg123, Linux's kTLS was broken. Which I see drewg123 also commenting in this thread. Is he the "Drew on my team" mentioned in the first video? Is he the speaker in the 2nd video? Idk https://news.ycombinator.com/item?id=28585008
I always use a configured!(F2) htop (not mentioned as well). Always enable PSI information in htop (some red hat systems I work with still don't offer them...).
If you have zfs enable those meters as well and htop has an io tab, use it!
My first command is always 'w'. And I always urge young engineers to do the same.
There is no shorter command to show uptime, load averages (1/5/15 minutes), logged in users. Essential for quick system health checks!
It should also be mentioned, Linux Load Average is a complex beast[1]. However, a general rule of thumb that works for most environments is:
You always want the load average to be less than the total number of CPU cores. If higher, you're likely experiencing a lot of waits and context switching.
[1] https://www.brendangregg.com/blog/2017-08-08/linux-load-aver...
On Linux this is not true, on an IO heavy system - with lots of synchronous I/Os done concurrently by many threads - your load average may be well over the number of CPUs, without having a CPU shortage. Say, you have 16 CPUs, load avg is 20, but only 10 threads out of 20 are in Runnable (R) mode on average, and the other 10 are in Uninterruptible sleep (D) mode. You don't have a CPU shortage in this case.
Note that synchronous I/O completion checks for previously submitted asynchronous I/Os (both with libaio and io_uring) do not contribute to system load as they sleep in the interruptible sleep (S) mode.
That's why I tend to break down the system load (demand) by the sleep type, system call and wchan/kernel stack location when possible. I've written about the techniques and one extreme scenario ("system load in thousands, little CPU usage") here:
https://tanelpoder.com/posts/high-system-load-low-cpu-utiliz...
Hey Tanel - I wanted to thank you for that blog post and psn tool - it recently helped me in a tricky performance investigation.
Glad to be helpful! :-)
The proper way is to have a idea of what it normally is before you need to troubleshoot issues.
What is a 'good load' depends on the application and how it works. Some servers something close to 0 is a good thing. Other servers a 10 or lower means something is seriously wrong.
Of course if you don't know what is a 'good' number or you are trying to optimize a application and looking for bottlenecks then it is time to reach for different tools.
Glances is nice. I think it is a clone of HP-UX Glance.
https://nicolargo.github.io/glances/
I have also hacked basic top to add database login details to server processes.
Me too! So much so that I add it to my .bashrc everywhere.
If you like this post, I would recommend “BPF Performance Tools” and “Systems Performance: Enterprise and the Cloud” by Brenden Gregg.
I have pulled out a few miracles using these tools (identifying kernel bottlenecks or profiling programs using ebpf) and it has been well worth the investment to read through the books.
Agreed, highly recommended reading. A slightly more up-to-date post of his which recommends tools in such situations is: https://www.brendangregg.com/blog/2024-03-24/linux-crisis-to...
Literally did miracles at my last job with the first book and that got me my current job, where I also did some impressive proving which libraries had what performance with it again... Seriously valuable stuff.
Yea it is kindof cheating. I was helping someone debug why their workload was soft locking. I ran the profiling tools and found that cgroup accounting for the workload was taking nearly all the cpu time on locks. From searches through linux git logs I found that cgroup accounting in older kernels had global locks. I saw that newer kernels didn’t have this, so we moved to a newer kernels and all the issues went away.
People thought I was a wizard lol.
I am curious, why don't you update regularly? (student here)
Kernel/distro upgrades can cause severe regressions. Generally the approach we do is have the service run a canary which runs on the newer kernel for a while to A/B test the upgrade. Generally we rely on the service owner to validate this A/B test as we don’t want to own making sure services are healthy on the new kernel. This means it is primarily on the service owner to look at the results of the A/B test to determine if the upgrade is ok.
I build a cloud (such as AWS) and we have many tenants running on the cloud. Much like AWS will not force upgrade AMIs, we will not force tenants to upgrade either.
Almost all of these have been replaced for me with below: https://developers.facebook.com/blog/post/2021/09/21/below-t...
It is excellent and contains most things you could need. Downside is that it isn't yet a standard tool so you need to get it installed across your fleet
Oh man nostalgia city. I vividly remember meeting atop time travel debugging at 3am in Menlo Park in 2012, wild times.
The iostat command has always been important to observe HDD/SDD latency numbers.
Especially SSDs are treated like magic storage devices with infinite IOPS at Planck-scale latency.
Until you discover that SSDs that can do 10GB/s don't do nearly so well (not even close) when you access them in a single thread with random IOPS, with queue depth of 1.
That's where you start down the eBPF rabbit hole with bcc/biolatency and other block device histogram tools. Further, the cache hit rate and block size behavior of the SSD/NVME drive can really affect things if, say, your autonomous vehicle logging service uses MCAP with a chunk size much smaller than a drive block... Ask me how I know
2015
Previous discussions: https://news.ycombinator.com/item?id=10654681 https://news.ycombinator.com/item?id=10652076
Yeah, I skipped the date and then saw Linux 3.13 in the examples.
Previously:
Linux Performance Analysis in 60,000 Milliseconds - https://news.ycombinator.com/item?id=10652076 - Nov 2015 (11 comments)
Linux Performance Analysis - https://news.ycombinator.com/item?id=10654681 - Dec 2015 (82 comments)
Linux Performance Analysis in 60k Milliseconds (2015) [pdf] - https://news.ycombinator.com/item?id=44070741 - May 2025 (1 comment)
After this article was written, `free -m` on many systems started to have an "available" column that shows the sum of reclaimable and free memory. It's nicer than the "-/+" section shown in this old article.
I quite like `iotop` as an alternative to iostat. https://linux.die.net/man/1/iotop
shameless plug: you can practice this in a free VM https://docs.sadservers.com/docs/scenario-guides/practical-l... (there's a typo there to keep you on your feet)
> At Netflix we have a massive EC2 Linux cloud
Wait a minute. I thought Netflix famously ran FreeBSD.
My understanding was their CDN ran on FreeBSD, but not their API servers. But I don't work for Netflix.
Your understanding is correct.
Why did they not choose to use it for both (or neither)? I.e., what reasons for using FreeBSD on CDN servers would not also apply to using them for API servers?
They are extremely different workloads so.. everything?
The CDN servers are basically appliances, and are often embedded in various data centers (includes those ran by ISP's) to aggressively cache content. They care about high throughput and run a single workload. Being able to fine tune the entire stack, right down to the TCP/IP implementation is very valuable in this case. Since they ship the hardware and software, they can tightly integrate the two.
By contrast, API workloads are very heterogeneous. I'd have to imagine the ability to run any standard Linux software there would also be a big plus. Linux clearly has much more vetting on cloud providers than FreeBSD as well.
Can't you fine tune linux as well? Does FreeBSD perform better somehow on a CDN workload? I find it difficult to imagine that the reason is performance. But I don't know what the reason is.
Netflix discusses their reasons starting at 18:20: https://www.youtube.com/watch?v=veQwkG0WdN8&t=18m20s
tl;dw: the performance, the efficiency of development, the community, FreeBSD is a complete operating system, the code base is smaller, the ports system, and the license.
and this video covers the optimizations Netflix has made to FreeBSD: https://www.youtube.com/watch?v=36qZYL5RlgY
Also potentially a reason: According to drewg123, Linux's kTLS was broken. Which I see drewg123 also commenting in this thread. Is he the "Drew on my team" mentioned in the first video? Is he the speaker in the 2nd video? Idk https://news.ycombinator.com/item?id=28585008
The CDN runs FreeBSD. Linux is used for nearly everything else.
Maybe I missed it, but checking available disk space is often a good step in diagnosing misbehaving systems.
it's 10 years later - what's the 60 second equivalent in 2025?
@yankcrime posted it above: https://www.brendangregg.com/blog/2024-03-24/linux-crisis-to...
PSI (pressure stall information) are missing.
I always use a configured!(F2) htop (not mentioned as well). Always enable PSI information in htop (some red hat systems I work with still don't offer them...).
If you have zfs enable those meters as well and htop has an io tab, use it!
Nice list. sar/sysstat is underrated imho.
Oh man. There's a blast from the past.
Today, you'd want something like:
Prometheus + Node Exporter [1]
[1] https://github.com/prometheus/node_exporter
he forgot about rusttop
I'm pretty sure that that didn't exist in 2015 ;)