Switzerland is lovely this time of year! Speaking of Switzerland, I have long been an avid reader of Storage Switzerland. I especially find George Crump’s articles to be not only informative but written in an engaging manner. The most recent example of this is his piece entitled, How to Judge Purpose Built Backup Appliance Performance. I could not agree more with the premise of this document and the points he makes.
I think his key point is that backup appliance performance needs to be viewed through an architectural lens, not just as numbers on a data sheet. As each vendor rushes to get their best number out as a headline, we do a disservice to customers who are just trying to understand how well a system will perform in their actual environment.
At ExaGrid, because our architecture is unique in the marketplace, we often find ourselves subject to misleading comparisons from other vendors, who mostly share the same type of architecture. As George rightfully points out: “Vendors usually don’t take into account architectural differences between products that are integral to how backup appliance performance is delivered. This makes getting to the point of an apples-to-apples comparison especially challenging with PBBA.”
Backup Appliance Performance Comparison
The key confusion points come from trying to compare monolithic controller-based architectures, like those of EMC, Quantum and HP to the second generation GRID-based architecture offered by ExaGrid. In the controller-based architectures, overall performance is provided by the single controller in a stack. Expansion occurs by adding a finite number of disk shelves to the controller. Each disk shelf added brings only capacity (i.e. workload) but no additional resources to increase backup appliance performance. The result is a growth in backup window as data grows and the eventual fork-lift upgrade when a controller hits its maximum capacity.
ExaGrid’s GRID architecture is very different. Each node is a full storage server. As such, it includes process, memory, network ports and capacity. As expansion occurs, full nodes are added, maintaining linear performance and avoiding costly fork-lift upgrades. The diagram below illustrates these key differences.

So, how do these differences influence backup appliance performance comparisons? Typically, the performance numbers for the controller architectures reflect the full capability of those controllers, regardless of price or how much capacity is initially involved. And, the performance number does not change as capacity is added to the system. So, a controller that can do 4 TB/hour will still perform at 4 TB/hour even after its disk capacity is doubled.
However, ExaGrid typically quotes a per-node performance number in our data sheets along with our GRID configurations. GRID configuration performance is a multiple of the per node performance. For example our EX13000E (node that would manage a 13 TB full backup) is capable of 2.4 TB/hour of ingest. But, our EX130-GRID (which is 10 EX13000Es in a single functioning system) is capable of 24 TB/hour of ingest. Meanwhile, the vendors with controller-based architectures will incorrectly claim that "ExaGrid’s fastest product" is 2.4 TB/hour. Let’s do a specific example.
Not to pick on EMC but let’s use a DD860 as a comparison point. The top stated performance number (with Boost) for the 860 is 9.8 TB/hour. The DD860 can handle around a 120 to 130 TB full backup. In most instances, EMC would compare that to a single EX13000E from ExaGrid and say “Look, we are 4 times as fast as their 2.4 TB/hour.” The missing part of the statement is that whereas the DD860 is built to handle up to 130 TB full backup, each EX13000E only manages up to a 13 TB full backup. It handles 1/10th the data while bearing 1/4th the ingest rate! A rightful comparison is between a DD860 and an EX130-GRID—both systems can handle a 130 TB full backup. In this case, the EX130-GRID is capable of 24 TB/hour of ingest. This is more than twice the backup appliance performance of the DD860. But only by factoring in the architectural differences can you get to an actual comparison.
The following table delivers a full set of example comparisons that should help folks sort this out.

For more information on backup appliance performance and architectural differences, check out our 5 Keys to Disk Backup Scalability whitepaper.