Latency has been scaling worse than almost any other factor for nearly the entire history of computing. The latency/throughput ratio for storage, networking, memory, cache and almost everything else has been rising continuously since the 70s and is likely to carry on rising.

This is going to have big consequences for the way that we design systems in the future. Transferring a large amount of data is going to be cheap and fast. Seeking, handshaking, back-and-forth and any other latency sensitive operations are going to be slow and expensive. This has already played a big role in algorithm design in HPC, and is going to start being felt to a much larger degree in the wider field over the next decade. As the latency/throughput ratio gets bigger, the tradeoffs behind optimal system design will change.

30ms for a Cray 1 was only 2400 clock cycles. 8ms for a modern CPU is around 30 million. That's a big change.

> 30ms for a Cray 1 was only 2400 clock cycles

2.4 million clock cycles. (According to Wikipedia, the Cray-1 had a 80 MHz clock, which you also mention in your other post.)

You are indeed correct. That's kind of how it worked in my head (about 12x change) but I didn't type it that way for some reason.

And buffer bloat is confusing the protocols in the process.

Average seek time across a disk which is much larger physically than the disks in use today.

It wasn't rare at all to find fair sized servos (as opposed to steppers in consumer grade stuff in the 80's) in those old disk pack units (the size of a washing machine).

You can't really compare 'common 7200 RPM drives' of today with a top-of-the-line medium from the 70's, physics didn't change at all in that time. That's why there are 'servo tracks' on the drive, they help with finding the right track (in a stepper scenario you don't actually need those, the stepper resolution defines where the tracks are).

Today enterprise level drives achieve < 4 ms average seek times at a price point which is a small fraction of what that drive cost in the 70's.

That's where the real improvement factor sits: performance (both in capacity, seek time and transfer rates) vs cost.

And that's many orders of magnitude.

Well, physics didn't change, but our understanding of it did. Modern read heads are based on the giant magnetoresistive effect which was only discovered in 1988. It's actually one of the fastest transitions from fundamental physics discovery to wide use (GMR read heads were widely used from 2001-2002 or so)

The actual mass of the heads and arms was an issue as well; you had to be careful not to seek back and forth at the natural resonance frequency of the cabinet or the "washing machines" would start walking across the floor.

A good SSD will have a seek time of something like 80 microseconds. Spinning mechanical platters are no longer the state of the art.

That's why they say that disk is the new tape. Releative to the sequential data rate, disk seeks are almost as expensive as a robot changing a tape reel used to be.

We aren't there yet, not by a long shot. Taking clock cycles as the base factor, a disk seek on a modern drive (8ms seek, 4GHz CPU) takes 32 million cycles. On the Cray1's 80MHz CPU, 32 million cycles is 400ms which would be really quick for a robot.

Tape seek times (for half a tape) are closer to 40s, which is still 100x more. We'll have to wait a while yet before disk is the new tape in terms of cost of seeks, even when comparing the 70s to today.

That's only considering latency. The cost consideration is a completely different story.

Why pleasantly surprising?

Because it is usually things like we have 1000x faster such and such and 100x smaller since such and such a time. This brings home that there were amazing technological advances even in the 70's worth noting. Sure we have SSD's now and so on. But, ya. Just kinda cool to see.