HDD Benchmark termonology

[Peaks]

Hey Guys,

I have been reading a lot of reviews for SSD's recently and have been a little confused about what some of the benchmark results actually represent. I think I know most of them but thought I would ask you guys see if someone can set me straight.

Sequential Read Performance: This is the amount of data that can be read sequentially (continuous block) from the drive, measured in MB/s.

Sequential Write Performance: This is the amount of data that can be written sequentially (continuous block) to the drive, measured in MB/s.

Sequential read/writes are important if you are using your drive for storage as this is the behaviour exhibited when copying MP3's/avi's etc.

Random Read Performance: This is the amount of data that can be read randomly (non continuous block) from the drive, measured in MB/s. Often measured in IOPs

Random Write Performance: This is the amount of data that can be written randomly (non continuous block) to the drive, measured in MB/s. Often measured in IOPs. Important if you are using your drive as a boot/application drive as this is the behaviour that will most likely be exhibited.

Read Latency: Delay between the read request and the drive performing the action. Latency for both sequential read and random read.

Write Latency: Delay between the write request and the drive performing the action. Latency for both sequential write and random write.

Read Access Time: Not sure, is this the same as read latency?

Write Access Time: Not sure is this the same as write latency?

Read Burst Rate: The highest speed in which data can be transferred form the drive to the operating system on a read request.

Write Burst Rate: The highest speed in which data can be transferred form the drive to the operating system on a write request.

Transfer rates: Not sure are these the same as read/write performances?

Throughput: Something like the average rate of data transfer. So again, is this the same as the read/write performances?

I may just be being stupid. I mainly look at benchmarks form HDTune and Everest so it could just be two different ways of saying the same thing.

 

[sub mesa]

Latency, Access Time, Seek Time and Service Time all are related:

Seek time: time to locate a sector on the mechanical disk
Access time: time to actually engage in a read/write operation (slightly higher than seek time+rotational delay)

Then you have controller latency, interface latency ("propagation delay") and all these combined make the Service Time aka latency. The service time it the time between issuing an I/O request from the host system, and actually receiving the data. So all latencies are included here and its the most relevant really.

The key issue here is that SSDs do not have "seek times" - just "access time" and its very low (0.08ms). So when doing random I/O an SSD can actually process alot of I/O operations where the HDD would be seeking 99% of the time and only 1% could it engage in actual reading/writing.

The write latency (or service time) was also a key issue in the JMicron JMF-602 SSD controller, causing "stuttering" because this simple controller doesn't do write buffering properly as it doesn't have a DRAM memory chip like HDDs do and modern SSDs as well.

The burst rates are used to measure the performance of the DRAM chip on HDDs/SSDs, so you won't be testing the mechanical part but the electronical part of a HDD. Not all applications do this right, however, but it should give you the interface bandwidth (i.e. with SATA/300 you should get close to 300MB/s burst speed).

Throughput and Transfer Rate are the same thing: the "data rate" if you will, expressed in MB/s for example. So when testing heavy random I/O on a HDD, your throughput will actually be less than 1MB/s. The "transaction rate" would be around 100-200 IOps, and the latencies would be between 8.0-22.5ms.

Basically, look at random read + random write performance especially for SSDs as the controller has alot of influence here. Some benchmarks do nice trace&replay benchmarks (like Intel iPeak suite; Rankdisk) but its not easy to perform or understand the I/O if you can't see exactly what its doing. With windows benchmarking all you see is numbers, not the mechanic behind it. I learned alot of monitoring my system with various utility during benchmarking, to understand the differences in scores i got, and it finally all made sense to me.

Good luck on your endeavors.

 

[sminlal]

For most intents and purposes, "latency" and "access time" can be treated as synonyms. There are some technical nits you can get into as Sub Mesa mentioned, but if a review or poster mentions one or the other of these terms and doesn't go into any more detail, you can assume they mean the same thing.

"Transfer rate" and "sequential read" / "sequential write" rates are also essentially the same thing.

"Throughput" is a vague term because it's completely dependent on the kind of I/O load being measured - is it sequential, random, or a mix of both? Single stream of requests or multiple concurrent requests? Without further qualification it's not a very meaningful number.

You said:

Random Write Performance: This is the amount of data that can be written randomly (non continuous block) to the drive, measured in MB/s. Often measured in IOPs. Important if you are using your drive as a boot/application drive as this is the behaviour that will most likely be exhibited.

...and in this I disagree. Random read performance, IMHO, is more indicative of OS boot performance since the OS disk is read from MUCH more often than it's written to.

 

[Peaks]

Thanks for the clarification guys, it is all clear no

@ sminlal : I agree with what you said about the random read being more important for OS boot. I was actually meant to put that statement as a side point suggesting that both random read and write are more important when concerned with an OS/application drive. Sorry for the confusion.