#1 PC Processing Bottleneck: One Queue Depth 4K Random Read
This is where the PC is busy trying to read random file locations quickly and can't catch up. Despite major technological improvements in other speed attributes, this bottleneck continues to plague today's PC in real-world performance tests. Even the world's fastest consumer NVMe drive only improves these test results by about 50%.
Throughout 2022 the industry has further obfuscated this weakness by replacing the "1 Queue Depth Random 4K Read" metric with IOPS (input output operations per second). IOPS reflects caching speed, not sustainable throughput. A high IOPS means you have a good cache and I/O onboard the device, which is great for bursts of performance, but it does not give you the reading you need for sustainable realworld performance. And so any manufacturer could slap a high end cache chip onto a substandard SSD and get a good IOPS rating. (This is very likely to fool the marketplace into thinking it's a good SSD, and any review site that participates will likely sell more ads to those manufacturers.)
Fortunately, for those looking for a more meaningful metric, you can still get the 1 queue depth random 4k read rating of a device by running 3rd party speed tests such as CrystalDiskMark. Just look for the number after the title: "RND4K Q1T1".
There are three primary PC bottlenecks:
- single queue depth random read
- sustained raw core speed
- virtual paging overflow
Virtual paging overflow used to be the most common bottleneck, because RAM used to be expensive, resulting in most PC's not having enough to adequately handle software and paging requirements. However, due to today's extremely low RAM pricing, this bottleneck is far less likely to surface, which is why we put it in 3rd place.
Here's the twist. These bottlenecks combine with caching deficiencies to produce the mechanisms behind planned obsolescence. No one can prove that it's intentional because the design deficiencies from different technologies combine to cause the problems. And it's generally easy enough to fix if you are aware of it.
The more powerful marketers do not want to touch this. It's making money the way it is. People believe their PC is going out-of-date, and that leads to another sale. As a result, all spotlights beam in other directions.
Thus why these marketers push market segmentation as hard as they can.
Today's biggest unspoken dilemma in PC marketing is that it is so common for people to think that their CPU is the bottleneck when it is more often than not the single queue random read bottleneck. And as a result, they go out and buy a new PC with a faster CPU. At first it's much faster. But then over time the new PC slows back down to the terrible performance level of their old PC. To the uninformed it then seems a good time to go buy a new one, again.
Did the PC go out of date?
What really happened is the single queue random read bottleneck kicked in gradually as their use of the PC crept past it's inadequate caching. This is the most significant mechanism causing people to mistakenly believe that their PC is going out-of-date.
|
A very common example of this is where they put a low-end Seagate hard drive into a lower priced PC, with a CPU that has an impressive specification, making it seem like a good value. Even though the CPU may be fast, the consumer's use of that PC will at some point expand beyond it's useful cache size. This causes that low-end hard drive to be accessed more and more often to read random locations, which is the slowest thing it can do, which then slows down everything the PC is doing.
Even the NVMe technology helps only a little. Unless it is large enough, the cache that comes with NVMe technology only temporarily fixes the bottleneck. This is easily proven through real-world performance tests.
|
A new PC will be fast only at first, because it takes time for this bottleneck to kick in. But just to be clear, what most old and new software requires, and what most PC's are terrible at, is good performing "1 queue depth 4K random read". New technologies come out each year making the other speed attributes faster, but this particularly important speed attribute continues to suffer, only progressing in tiny increments every few years, and therefore remains a serious performance bottleneck, even in brand new PC's.
If your software is too slow, and you wish to speed it up with a faster PC, then this issue should be very important to you.
Even the fastest Samsung NVMe drive only speeds up 1 queue depth 4K random read by about 50%. All the others? Hardly at all.
One solution is the Intel Optane SSD (not the accelerator, the SSD). However, it is quite expensive, and therefore only suitable to high-end servers.
The only very good and very affordable solution is a properly designed disk-cache, which is ordinarily the operating system's responsibility.
It potentially accelerates single queue 4K random read by 60x, but it requires more RAM to make this work. RAM is now very cheap, making this the very best solution.
And then there's the fact that Microsoft defaults their File System Cache to a dynamic allocation, rather than fixing its size to an optimal level. This decision was clearly left over from the expensive RAM era, to better handle bottleneck #3 (see above). In a RAM rich PC it's better to lock in the File System Cache to just the right fixed size, to prevent an almost inevitable tug-of-war that will result in one of the bottlenecks resurfacing despite having plenty of RAM.
If you don't lock the File System Cache size, then some software will likely hog its memory to the point of bringing back bottleneck #1. In these cases you'd rather get a low memory error from the software to know to add more RAM. But Microsoft is not facing this specific point (likely for reasons explained above). Instead, leaving the allocation logic the way it is results in one bottleneck or the other losing the tug-of-war. And you get the idea that your PC is going out of date.
If you don't know what to do, the safest bet is to set both the minimum and maximimum cache size to about 1/2 of your total RAM, and leave it there. There are all sorts of possible issues and this setting is the most likely to resolve them.
You can calculate the right amount of RAM to buy as follows... First add together the RAM requirements for each software package that you intend to use simultaneously. Then add together the sizes of their active resource files. Then if you are a light-weight user add another 2GB, for medium-weight add 5GB, and for heavy-weight add another 10GB. Then round up the total to the nearest power of 2 (4GB, 8GB, 16GB, 32GB, or 64GB). This is a very good guess as to the amount of RAM that you need.
Between choosing the right amount of RAM, and configuring the File System Cache correctly, and adding the right NVMe product, you can eliminate the single queue depth 4K random read bottleneck. And that then prevents your PC from getting slower over time (assuming you perform all the normal disk and OS clean-up tasks).
We install this fix for you on each new Silver Bullet PC.
Your PC will no longer creep past a nebulous limitation, unwittingly turning on a bottleneck that had no reason to be there in the first place.
|