This blog post was originally written by Rajkumar MR and Chuck Farah.

Comments are welcome! To suggest a blog topic or make other comments, contact WinServerBlogs@dell.com.

To read more about Windows Server 2012 and Dell, go to the Windows Server 2012 page on Dell TechCenter.

 

Offloaded Data Transfer Overview

Window Server 2012 and EqualLogic follow SCSI command standards to start a copy request through a series of tokenized responses. The term in this case means that less data packages will be required to communicate the reads and writes from source and destination volumes.  By communicating directly to the EqualLogic array the typical LAN network and Host resource consumption are minimized.

ODX adopted the synchronous offload SCSI command to support advanced MPIO, failover cluster support and avoid SCSI command time-out.

In our earlier ODX blog, we discussed the functionality of ODX features. In that post, both Windows server 2012 as well as Dell Equallogic firmware was in beta stages. As we expected there has been a huge performance gain in the latest release of both the products.

 

Figure 1 ODX vs. STD Copy data flow – Left side represents ODX enabled firmware; Right side represents non-ODX enabled firmware

Note: The standard copy consumes CPU, memory and network on the SAN attached servers, while the Offloaded Data Transfer has little impact on these resources.

In this blog, we will compare the difference as well as some of the best practices from our testing.

Performance Results

The table shown below compares the data transfer performance of Windows Server 2012 with ODX enabled and Windows Server 2012 with ODX disabled (a registry entry was changed to disable this feature).  A copy operation of 11 GB data from one volume within an ODX enabled PS Series Group took on average 20 seconds compared to 108 seconds without ODX. For a use case were the Client is separated by long distances or constrained by network or CPU the impact would be even more impressive. Consider a situation where a user needs to quickly move a virtual machine file while 2000 miles away from the data center. ODX would enable this task which would be almost impractical due to bandwidth and latency constraints using standard copy.

 

Windows Server 2012 Standard Copy

Windows Server 2012 ODX-Enabled

Improvement

CPU Utilization

42% (2.3GHz)

10% (2.3GHz)

~ 76%

Network Utilization

37%~100% (1GigE)

0%~1% (1GigE)

~ 99%

Copy time of 11GB

 

108 seconds

20 seconds

~ 5x

Speed

104MB/s

630 MB/s

~ 5x

Claims are:

  • Up to 76% improvement in CPU utilization *
  • Up to 99% improved network utilization *
  • Up to 5 times improved copy time *
  • Up to 5 times faster data transfer *

*Results based on July 2012 internal Dell testing using Windows Server 2012 ODX Enabled compared to Windows Server 2012 with-out ODX enabled or Standard copy.

This performance results clearly demonstrates the possible improvements for compute resource efficiencies.

 

Note: These tests were done in isolation. There was no additional workload or application process running at the time of tests. If there had been additional workload or application processes running, then there would typically be a more pronounced difference. In short, the more pressed for network bandwidth and CPU cycles, the greater the benefit ODX.

 

Offloaded Data Transfer (ODX) Performance Considerations:

Network between Arrays:

For volumes that have large amounts of data transferring between members (separate pools or same pool) the link between the arrays may be the limiter. For instance if you have members with gigabit Ethernet interfaces and your volume spans more than one array the overall bandwidth would be limited by that network capability within that group or pool.

For greatest efficiencies consider 10GigE arrays in the same group and reading from one array or pool and writing to the other.  This will imply two separate arrays or Pools at the minimum. However transfers still can occur within a single member or Pool.

Disk Technologies

NLSAS is often a very cost effective technology with good performance characteristics, particularly when large sequential data such as ODX are involved. On the other end of the spectrum SSD technology is optimum in very high random read environments. SAS 15K RPM or 10K RPM disks will provide the most predictable performance for a multitude of read/write patterns, block sizes and latency.

Additional I/O Activity

The ability for the PS Series SAN to accommodate multiple I/O loads at once should be considered to set expectations for the copy of data using technologies like ODX. While the ODX copy will generally still perform more efficiently even when there are other workloads present the speed of completion will be dependent on array resources available to perform the reads and writes.

 

Conclusion

Offloaded Data Transfer with Windows Server 2012 and PS Series firmware provides accelerated data movement. The capabilities of this technology help open up new business use cases for PS Series SANs in Microsoft environments.  

For more detailed information regarding Windows Server 2012, go to the whitepaper Understanding Microsoft Offloaded Data Transfer on EqualLogic arrays.