Making sense of new CrystalDiskMark?

[Quattro Pete]

A few weeks ago the author released a new version (5.0.2) of this disk test utility. It is now using some new benchmark methodology, so the results are not comparable to version 3.0.x. I am just trying to understand what these new tests exactly are. There doesn't seem to be any explanation there on this so called "Help" page...

http://crystalmark.info/software/CrystalDiskMark/manual-en/

Let's say the test is 4K Q32T1. I understand the "4K" part, but what are queues and threads? How do they affect disk read/write speed? Is there a standard number of queues and threads that one should select when performing these tests to make them representative? Would you use different values for testing an SSD vs HDD vs flash memory?

Thanks!

 

[C1KLTX]

QD 32 or Queue Depth 32 is basically how many requests are made.

 

[Quattro Pete]

You mean like trying to write/read 32 of those 4K blocks at once?

What are threads then?

 

[razel]

The last two results seq (Sequential) and 4k hasn't changed.

SeqQ32T1 = Sequential, 32 Queue, 1 Thread, Maximum speed with 32 disk requests, one CPU instruction or one at a time. Trying to measure the maximum speed of the interface minus any CPU help.

4KQ32T8 = 4k, 32 Queue, 8 Thread, Means 4k (most common smallest block of data) 32 requests, 8 simultaneous. IMO is trying to measure the advantage of going NVMe.

Standard queue for home/office usage. With SSDs it's almost always below 1 queue. Only when you do windows udpates/spyware scan will you see 1-4 queue. 4+ is pretty rare. I've seen 8 queue only when I directly stress the disk and is beyond everyday use.

As for different values, there are only the obvious ones. HDDs have slow random 4k reads usually under 1. $10 Flash drives have embarrassing random 4k writes WELL under 1. SATA SSDs will max out sequentials for their maximum rates.

NVMe... wow. It's amazing. Early benchmarks from pcper showed that the programming of benchmarks needed to be changed since CPU speeds began to affect benchmarks. The issue was benchmarks were nearly all 1 thread. So here we are. CDM with 8 threads. These are just 1st gen NVMe products. Can't wait to see the years ahead. I think all the data it's pushing (data = electricity) we may need to begin putting heatsinks on the NAND chips. Seriously.

 

[Quattro Pete]

Thanks.

 

[Rand]

NVMe is PCI express interface SSD.

What they really need to do is rethink the whole way SSD store data. They shouldn't be locked into place using data "strategies" designed for HDD.

That being said for the residential market. SSD are good enough for years to come.

Going from ~100 IOPS(HDD) to many thousands in normal operation is quite sufficient.


But there are still MASSIVE gains to be had when using Solid state storage for certain applications such as Database.. esp in commercial/business applications.

 

[razel]

Just saw a review of the brand new CPU/chipset from Intel. The i7-6000's and their matching chipset. ALREADY with 2 Intel 750's NVMe SSDs Raided... it has already maxed out beyond what the CPU can handle on it's internal PCIe connection. Max throughput is 4000MB/s. Raided NVMe... about 3500MB/s. When you add overhead, it pretty much has topped what the latest CPU from Intel can handle.

Perhaps with an enterprise CPU with more PCIe lanes you can get more but wow. That is faster than DDR2 PC3200 3200MB/s memory. Of course latency 2 to 3 times order of magnitude faster with RAM but still. I shake my head in NVMe amazement. It has a VERY bright future.

 

[javacontour]

NVMe is part of our Engineered Solutions such as Exadata.

http://www.oracle.com/technetwork/server...ure-2328157.pdf

Seldom was the CPU a bottleneck. Things could be changing as storage keeps getting faster and closer to the CPU.

Originally Posted By: razel
Just saw a review of the brand new CPU/chipset from Intel. The i7-6000's and their matching chipset. ALREADY with 2 Intel 750's NVMe SSDs Raided... it has already maxed out beyond what the CPU can handle on it's internal PCIe connection. Max throughput is 4000MB/s. Raided NVMe... about 3500MB/s. When you add overhead, it pretty much has topped what the latest CPU from Intel can handle.

Perhaps with an enterprise CPU with more PCIe lanes you can get more but wow. That is faster than DDR2 PC3200 3200MB/s memory. Of course latency 2 to 3 times order of magnitude faster with RAM but still. I shake my head in NVMe amazement. It has a VERY bright future.