One thing I find interesting is how much tuning of config variables is necessary to get "good" application-level performance.
C-states, P-states, turbo, general prefetch, "buddy" prefetch, etc., at the CPU level, GC configs, etc., at the runtime level, retry policy, thread pools
C-states, P-states, turbo, general prefetch, "buddy" prefetch, etc., at the CPU level, GC configs, etc., at the runtime level, retry policy, thread pools
https://twitter.com/rksio/status/1483124351276765184
, etc., at the application level; etc.
A doc I saw recently advocated for Go over Java because it has fewer GC config cars because incorrect GC configs are a major cause of incidents. But not having knobs doesn't solve the problem; it's equivalent to picking an arbitrary config.
A doc I saw recently advocated for Go over Java because it has fewer GC config cars because incorrect GC configs are a major cause of incidents. But not having knobs doesn't solve the problem; it's equivalent to picking an arbitrary config.
Some large companies have an automatic config tuning framework, but very few companies are large enough to want to have a full-stack optimization setup where apps get run on specific hosts that have machine-level configs that are compatible with the application.
Although there are efficiency gains from using "the cloud", not being able to tune host-level stuff is a large and measurable efficiency loss.
Younger companies with the scale/capital to operate their own DCs efficiently (e.g., Stripe) seem to have no desire to do so.
Younger companies with the scale/capital to operate their own DCs efficiently (e.g., Stripe) seem to have no desire to do so.
[Back of the envelope numbers]
People are outraged crypto uses 0.5% of world electricity. But 10% of the world's electricity is used by computers and I've never seen a system at a < $900B company where I couldn't get a 10% efficiency win by changing config (frequently >= 50%).
People are outraged crypto uses 0.5% of world electricity. But 10% of the world's electricity is used by computers and I've never seen a system at a < $900B company where I couldn't get a 10% efficiency win by changing config (frequently >= 50%).
I'm outraged by general inefficiency the same way some people are outraged by bitcoin power consumption. This seems to be uncommon, but I don't understand why since, everywhere I look, I see very easy to fix waste.
https://twitter.com/danluu/status/1431205440411639815
Just for example it's pretty common to see applications with ~50% waste due to lack of thread pool size tuning combined with the use of CFS bandwidth control (by far the most commonly used Linux container-level CPU isolation mechanism).
https://twitter.com/danluu/status/1472142011918471170
This isn't unique to tech; a competent engineer in any field can find huge waste everywhere they look.
This isn't really talked about much by lay people, the media, etc., because, e.g., disposable plastic packaging is more legible than this thread, but maybe that should change?
This isn't really talked about much by lay people, the media, etc., because, e.g., disposable plastic packaging is more legible than this thread, but maybe that should change?
• • •
Missing some Tweet in this thread? You can try to
force a refresh
