Every once and again, I like to go back to old articles I have written one, maybe two years ago to see what was the exciting news of that time and gain some perspective on how far we've come, or not. It just so happens that at this time last year, I was writing about the Direct Assess File System Collaborative.
The mission of the DAFS Collaborative is to increase remote file system performance by using remote direct memory access techniques. Yet another method for increasing network-based data transfers (see my recent articles on Transport Offload Engines). So it seems we, as an industry, are still concerned about network data performance and we still don't have all the answers.
So what are the options available to customers today for increasing the performance of their IP-based storage networks?
Well, part of the answer is faster pipes, so having 10/100 Base-T Gigabit Ethernet would certainly help.
Once you leave the physical pipe, you run into nebulous land where hardware and software join to create solutions. One option, as described above, is to use TOEs. As previous articles outlined, customers can gain performance on specific platforms but with the market still in the early stages, it wouldn't be prudent to try to implement TOEs on every platform in a company, or even data center.
Another solution, which is very new, that sits between the wire and the server, is an InfiniBand switch/router. These switches take IP traffic and depending on the implementation, terminate the IP stack within the switch, translate to InfiniBand then use RDMA methods to move the data into memory. While this is an efficient way to improve performance, right now, it's an expensive way as well.
Of course, there is DAFS, which quickly moves data off the wire and into memory. However, for applications to use DAFS, modifications must be made. If, as a customer, you are in control of the application and experiencing performance issues, this may be a good option to investigate. However, the rest of us, who relay on middleware applications, may be waiting quite a while for performance improvements using DAFS.
Along similar lines of DAFS, the network file system folks are trying to get RDMA NFS into the next specification. This basically does the same thing NFS does so software doesn't require changes, but the underlying calls can use RDMA methods or do it the way it's done now. This solution is still - in the labs.
The final place where performance can be gained is within the system itself. The I/O channel has an effect on the performance, depending on whether there is a bus or channel architecture, the memory and processor speeds make a difference and, of course, the operating system implementation can make a difference. While these are not things end-user customers are likely to have a direct effect on, you exert great influence over your vendors. Asking questions and demanding capabilities is your leverage.
I imagine if I had been writing these newsletters for the last 10 years, I could have written about similar issues. Some of the technologies coming out today are helping the problems today but as a whole, we tend to consume all resources given us. So next year, it will be interesting to see what the performance expectations will be and what new technologies or thoughts are in the forefront to solve the performance bottlenecks.