Block sizes

classic Classic list List threaded Threaded
3 messages Options
Reply | Threaded
Open this post in threaded view
|

Block sizes

Jori Gielis
Hello,

Me and my colleague are ICT students at Thomas More Geel in Belgium.
This year we are doing a 3 months internship at the company Valcosoft.
The company has given us a subject, this subject is iSCSI SAN. 
As SAN software are we using Open-E.
We are testing the performance of several systems. 
If we test our system with 1Gbit connection we get speed of 9,60MB/s at block size 4KB. At a block size of 4MB we get a speed of 106MB/s. If we use the same system except we replace the 1Gbit card with 10Gbit card. At a block size of 4MB we get  a speed of 212MB/s. What part of our system is limiting the speed at 4KB? Why are we able to get better performance at large block sizes?

Hardware specs:

Model

Cisco UCS C210 m2

Operation system

Open-e DSS  v.7.00 up12 10529

CPU

Intel Xeon E5606 2.13 GHz

Motherboard

Intel 5520 chipset

RAID controller

onboard SATA RAID 0/1 controller

Interface

SATA2

RAID

FakeRAID

RAID levels

0 en 1

Chipset

intel ICH10R

Memory

2 x Cisco Systems N01-M304GB1-L

Capacity

4GB

speed

DDR3 1333 MHz



startup device

USB stick USB2.0


Solid State Drive

Samsung 840 Pro series MZ-7PD128

Type

MLC

Capacity

128GB

Cache

256MB

Interface

SATA3 6Gb/s

speed

 

read (sequential)

530MB/s

write (sequential)

390MB/s

read (random en blokgrootte 4kB)

97000 IOPS

speed in bytes = IOPS * block size   397,312MB/s = 97000 * 4096 bytes

write (random and blokgrootte 4kB)

90000 IOPS

speed in bytes = IOPS * block size 368,64MB/s = 90000 * 4096 bytes

Network adapter 10Gbit

Intel X540-t2 ethernet converged network adapter

speed

10Gbit



We hope you can help us.


Thanks in advance.


 

Yours faithfully


 

Jonas Deckers and Jori Gielis




------------------------------------------------------------------------------
"Accelerate Dev Cycles with Automated Cross-Browser Testing - For FREE
Instantly run your Selenium tests across 300+ browser/OS combos.
Get unparalleled scalability from the best Selenium testing platform available
Simple to use. Nothing to install. Get started now for free."
http://p.sf.net/sfu/SauceLabs
_______________________________________________
Iometer-user mailing list
[hidden email]
https://lists.sourceforge.net/lists/listinfo/iometer-user
Reply | Threaded
Open this post in threaded view
|

Re: Block sizes

Carl Zwanzig-2

Hi Jori,

 

A couple of things-

I assume the specs below on the Samsung drive are from Samsung, not from measurements. I really doubt that drive would make more than 40k IOPS in real usage.

Benchmark the test drive directly in the client system, then you’ll have a better idea of its reality.

Don’t bother with 1g Ethernet for performance test if 10g is available.

 

To examine the question  ”What part of our system is limiting the speed at 4KB?”

What io queue depth are you using? I’d go to at least 32.

What is the storage network MTU? If the 1g net is 1500 and the 10g is 9000, you’ll get very different speeds.

 

Let us know what you find.

 

z!

 

 

From: Jori Gielis [mailto:[hidden email]]
Sent: Thursday, May 22, 2014 11:22 AM
To: [hidden email]
Subject: [Iometer-user] Block sizes

 

Hello,

 

Me and my colleague are ICT students at Thomas More Geel in Belgium.
This year we are doing a 3 months internship at the company Valcosoft.
The company has given us a subject, this subject is iSCSI SAN. 

As SAN software are we using Open-E.

We are testing the performance of several systems. 

If we test our system with 1Gbit connection we get speed of 9,60MB/s at block size 4KB. At a block size of 4MB we get a speed of 106MB/s. If we use the same system except we replace the 1Gbit card with 10Gbit card. At a block size of 4MB we get  a speed of 212MB/s. What part of our system is limiting the speed at 4KB? Why are we able to get better performance at large block sizes?

 

Hardware specs:

Model

Cisco UCS C210 m2

Operation system

Open-e DSS  v.7.00 up12 10529

CPU

Intel Xeon E5606 2.13 GHz

Motherboard

Intel 5520 chipset

RAID controller

onboard SATA RAID 0/1 controller

Interface

SATA2

RAID

FakeRAID

RAID levels

0 en 1

Chipset

intel ICH10R

Memory

2 x Cisco Systems N01-M304GB1-L

Capacity

4GB

speed

DDR3 1333 MHz

startup device

USB stick USB2.0

 

Solid State Drive

Samsung 840 Pro series MZ-7PD128

Type

MLC

Capacity

128GB

Cache

256MB

Interface

SATA3 6Gb/s

speed

 

read (sequential)

530MB/s

write (sequential)

390MB/s

read (random en blokgrootte 4kB)

97000 IOPS

speed in bytes = IOPS * block size   397,312MB/s = 97000 * 4096 bytes

write (random and blokgrootte 4kB)

90000 IOPS

speed in bytes = IOPS * block size 368,64MB/s = 90000 * 4096 bytes

 

Network adapter 10Gbit

Intel X540-t2 ethernet converged network adapter

speed

10Gbit

 

We hope you can help us.

 

Thanks in advance.

 

 

Yours faithfully

 

 

Jonas Deckers and Jori Gielis

 

 


------------------------------------------------------------------------------
"Accelerate Dev Cycles with Automated Cross-Browser Testing - For FREE
Instantly run your Selenium tests across 300+ browser/OS combos.
Get unparalleled scalability from the best Selenium testing platform available
Simple to use. Nothing to install. Get started now for free."
http://p.sf.net/sfu/SauceLabs
_______________________________________________
Iometer-user mailing list
[hidden email]
https://lists.sourceforge.net/lists/listinfo/iometer-user
Reply | Threaded
Open this post in threaded view
|

Re: Block sizes

Carl Zwanzig-2

[please keep replies on the list]

 

> We are using one 1 IO queue depth. What does this value mean?

IO queue depth is the number of operations that may be initiated before waiting for any of them actually complete (sort of like the tcp sliding window).

 

> We are using 1500 bytes for both connections. 

You will get higher throughput with a larger MTU because there will be less segmentation of the data blocks. (Some systems default 10g Ethernet to MTU=9000, some don’t.)

 

> Why do we get more performance at bigger blocks? 

Isn’t that part of what you’re researching? Look to the overhead involved with each operation, with iscsi, with tcp & ip, and with Ethernet frames; that will get you part of the way.

 

z!

 


------------------------------------------------------------------------------
"Accelerate Dev Cycles with Automated Cross-Browser Testing - For FREE
Instantly run your Selenium tests across 300+ browser/OS combos.
Get unparalleled scalability from the best Selenium testing platform available
Simple to use. Nothing to install. Get started now for free."
http://p.sf.net/sfu/SauceLabs
_______________________________________________
Iometer-user mailing list
[hidden email]
https://lists.sourceforge.net/lists/listinfo/iometer-user