Question What are "threads" in SSD benchmarks

[tommo1982]

I can't find a decent explanation. What are the 'threads' in Q32T1? How does it affect performance?

 

[IntelUser2000]

Those are just CPU threads. The higher thread counts are more demanding to rest of the system, including the I/O such as memory and storage.

 

[tommo1982]

A for you Sir.

 

[nosirrahx]

A good analogy when it comes to queue depth and threads would go something like this.

You have people and each person has a list of requests for information. The number of people are the threads and their list of requests is the queue depth.

4KQ4T8 would mean there is 8 people asking for 4K of data 4 pieces at a time.

When it comes to SSD speed 4KQ1T1 is a good metric as it puts the drive in a position where it is very weak.

The 4K part is a small bit of data. Storage works faster the fewer times it needs to jump around looking for the next chunk of data so these small 4K chunks are not ideal.

Q1 and T1 means that nothing overlapping is happening. 1 person is requesting 1 4K chunk and then repeating.

If a SSD is very good at this metric, it will 'feel' very fast as an OS/apps drive.

 

[tommo1982]

Even more to you good Sir.

Anyway, the number of active threads is limited to CPU threads, as stated above?

 

[IntelUser2000]

Well, you can increase it, but there would be no point.

 

[nosirrahx]

Even more to you good Sir.

Anyway, the number of active threads is limited to CPU threads, as stated above?

Threads here isn't all they related to the CPU as under typical use cases your CPU wont ever be using loads of threads to do disk stuff.

High thread count comes into play when you have a bunch of VMs all talking to the same disk cluster. In this case the disk access threads can be huge as all of those users are working with the same data at the same time.

This is why we talk about 4KQ1T1 so much, its the metric that matter most for almost all users.

 

[IntelUser2000]

This is why we talk about 4KQ1T1 so much, its the metric that matter most for almost all users.

Ehh, not really. Up to queue depth of 4 is quite commonly used. Sequential throughput is quite important which is why Optane isn't dominant in consumer workloads.

Main storage used in modern computer systems are what's referred to as "cold" storage. The data is rarely used. DRAM is "hot" storage, because the CPU needs to frequently access it. That's why with DRAM, the transfer speed matters, and impacts your frame rates, and makes your CPU compute faster. With faster SSDs it does no such thing.

 

[nosirrahx]

Ehh, not really. Up to queue depth of 4 is quite commonly used. Sequential throughput is quite important which is why Optane isn't dominant in consumer workloads.

You are missing the point. 4KQ1T1 is a good metric for how things perform when at their worst. If this metric is good, anything more complex will be great.

You are 100% wrong about Optane. If it wasn't for everything it gets wrong (price, complexity, annoying form factors, capacity....) People would actually prefer it. I build every system now with Optane + SATA SSD, performance is better than NVMe for common PC user workloads by quite a bit.

This is what you get if you are willing to deal with all of the BS (you mentioned Q4 so take note of that score):

 

[IntelUser2000]

You are 100% wrong about Optane.

What am I wrong about Optane? That it doesn't perform worth the price increase?

I like it, which is why I created a dedicated thread for it. But just like Turing cards, it has the problem of being too expensive. Anand had a saying "there's no bad products, just bad pricing".

Optane has a further problem. They are not really making money on it.

 

[nosirrahx]

What am I wrong about Optane?

Sequential throughput is quite important which is why Optane isn't dominant in consumer workloads

I have systems with the 950 pro, 960 pro and 970 pro, that glorious sequential speed is more or less a non-factor in most operations.

The only time I have seen sequential speed really crush it is when moving very large files from 1 fast drive to another fast drive. If you were moving a collection of high quality rips from 1 970 pro to a different 970 pro the sequential speed would really shine. How often does the average used do something like that?

In general Optane is a cool tech toy that represents what future storage will be like. Its a fun gadget and can be used in high end builds but in general it is many miles from consumer ready. Its too bad because the stuff is pretty awesome.

 

[IntelUser2000]

That's not wrong.

The reviews show this. Optane is substantially better in low queue depth, it does not have to do garbage collection, it does not slow down after fill up, and it does not slow down with large file deletes.

Yet the competition is still better at some things than others. You can't brush sequential performance away because some people do care about file transfers, making Optane not a uniformly faster drive.

 

[nosirrahx]

That's not wrong.

The reviews show this. Optane is substantially better in low queue depth, it does not have to do garbage collection, it does not slow down after fill up, and it does not slow down with large file deletes.

Yet the competition is still better at some things than others. You can't brush sequential performance away because some people do care about file transfers, making Optane not a uniformly faster drive.

If I took 100 random users and monitored their disk access, sequential movement of huge files between 2 drives is not going to be super common.

Typically people that have a massive collection of rips and ISOs store them on more traditional HDDs due to the insane price and complexity of multiple TB solid state storage.

Ironically the best way to currently do 8TB of SSD storage is to cache those 8TB Micron SATA SSDs to an 800P. This is about $1000 cheaper than 4 2TB 970 EVOs in RAID 0 and less complex.

Of course Intel in their infinite wisdom does not officially support this even though Optane is literally a Micron and Intel project.