Welcome Guest | Sign In
E-Commerce Times TechNewsWorld CRM Buyer LinuxInsider

E-Commerce Times Talkback

ECT News Community   »   E-Commerce Times Talkback   »   Re: SANity Check

Re: SANity Check
Posted by: Jim Metzler 2004-01-21 17:22:05
See Full Story

Along with any new technology come large doses of hyperbole and vague promises. With storage-area networks, the promises included better ROI through storage consolidation and increased business productivity through collaboration. Now that SANs have been in place for a while, it's possible to step back, assess the scope of SAN deployment and gather some insight into how SANs are being designed to support business applications.

Tools do exist, but the solution starts with architecture modeling
Posted by: lbudnik 2004-01-21 17:22:37 In reply to: Jim Metzler
Application Brownout Analysis
The concept of an application brownout is certainly difficult to isolate because of the many layers of virtualization at play. From an ITIL resolution perspective, this is a capacity problem. These problems begin during the architecture phase when expected service levels and performance expectations are set. The balance between cost and capacity takes hold resulting in something less than the expectation. While the article suggests that there are no monitoring tools, SNMP and APIs are always with us. There remains no substitute for modeling.
The move to consolidated storage infrastructures accessed via fabric provided a huge performance improvement. Utilization of disk space increased by using storage arrays, port count through using fabrics and staff. We played the odds and won because most applications cannot drive the channel or the disk. However, some applications natively can drive the channel all the time and most systems have times, typically during backup, when they can drive the channel to maximum utilization.
At the array level, we spread multiple workloads among virtualized disks. We gain processor time through a transfer of protection to the array. We gain higher throughput than is possible from a single disk by spreading each workload across many spindles. Yet most do not manage the overall load on the spindles, much less understand the impact of a single workload yet we expect it to run as if there was only one workload.
The concept of "Access Density = (total IOPS of all workloads per spindle / GB per spindle)" is often neglected as drive sizes increase we are unwilling to abandon storage space in the interest of performance. However, you can measure the spindle, LUN or hyper-volume with today's tools. In most arrays, you can get the data via SNMP or APIs and in the future via the SNIA SMI-S data structures. Of course, there are many other data points you can gather within the arrays.
At the network level there is a maturity evolving that allows acceptance of core-edge designs (although some may never change.) Maturity plus new security offerings are enabling larger port counts per fabric. Yet, the rules of thumb for fan-in and fan-out ratios used during design do not represent real workload modeling. Apart from flow, understand the request sizes and quantities as FC and arrays work better with large request sizes and really poor on small requests (you can actually max out the links request rate and only send 2M down a 200M pipe.) It is all about IO. Again, the switches are SNMP enabled and you can collect and analyze the data.
You might think that if I were a doctor that I would suggest that "this is all in your head and that there is nothing wrong" with your storage infrastructure, but that is not the case. The operating systems, fabric and arrays all have particular quirks to manage by rule of thumb. My favorites include anything that makes the channel synchronous. For this reason, if possible I isolate tape and disk HBAs, keep file systems under 100% utilization, etc. I choose to plan the access of large, data intensive systems and group their I/O with isolation mixed with failure coverage. I still place data for the top three applications in the infrastructure. I accept that the rules of physics continue to apply and that you must know the capability of your components.
As usual, we all have time to fix the problem but never time to model so that the problem does not occur.
Jump to:
Which type of online advertising is most likely to attract your favorable attention?
Straightforward display ads
Ads based on my interests
Informational articles on products/services
Video ads
Ads designed to grab my attention, e.g. pop-ups, autoplay
None -- I avoid all online ads