Suicide scheduling has never been officially supported, and in fact was explicitly stated that at some point they might choose to limit it to some period of time, such as every 5 minutes minimum:
While we’re on the topic of chaining, I’d like to describe another
disturbing trend, one that I call “suicidal scheduling”. Suicidal
scheduling is where a scheduled job enqueues another scheduled job and
then aborts itself. This allows these jobs to be chained.
Try stopping this job if you want to. You can’t. It goes too quickly
to be aborted. By the time you have found the job ID, the job has
already run, enqueued another, and aborted itself. It’s like a virus,
only you wrote it yourself, instead of unintentionally downloading it
with the FREE NO ADS version of Candy Crush Saga.
We will not block this approach, since it functions today. We will,
however, slow it down. Any call to System.schedule inside of a
scheduled job will run no sooner than five minutes later. Ostensibly,
the job is a “scheduled job”, which means it is going to repeat itself
lots of times, so five minutes shouldn’t be a major issue. If it is
set up simply to kill itself, throughput is going to degrade. That is
good for everyone else on the pod.
They've been threatening this for three years now. I even asked a question about this.
If you really want to improve your performance, consider scheduling 30 jobs that run hourly and are staggered every 2 minutes. Or maybe it's time to consider a new approach, like using Heroku or AWS EB, which gives you more control over the processing time in exchange for API calls.
The system has no guarantee about how long it will take asynchronous code to run at all. It could be exactly on the second, or it could be delayed by hours. The system runs jobs as fast as it can given the resources it has available. Most likely, your org has been doing too much and so the system is giving time to other orgs.
Also keep in mind that there is a 250,000/day limit on asynchronous calls, spread between batch executions, queueable, and scheduleable. If you're brushing up against that limit, that would also cause your jobs to be delayed.