Is There Hype about Cloud Performance?
I keep reading articles about how Cloud Computing is "high performance," especially from vendors and their partners who are eager to sell it. Unfortunately, these discussions don't typically define what "high performance" is. For example, today's excellent article on Cloud at Network World gave performance only a cursory discussion at the end. Many people assume that today's microprocessors are so powerful that today's Clouds can handle any application, but my experience with enterprise applications is that overall performance depends on a lot more than just CPU speed or RAM size, and is very dependent on exactly what the demands of the application are (and therefore the actual design of the application as well.)
What I've seen is that discussion of cloud computing performance often avoids difficult topics that cloud vendors would rather not discuss because it might tarnish the current impression that Cloud is high performance and suitable to every need. While Cloud deployment can benefit a majority of applications, it isn't suitable to all though I believe it will be eventually. Let's take a look at some of the performance factors affecting the Cloud.
Latency (between the user and the Cloud or between elements in the Cloud) is often the only performance factor that is discussed, and it is important: we get customers from Amazon that talk of seconds of latency in some cases. But as many in the industry are quick to point out, latency is dependent to a significant degree on the "last mile" of network leading to the consumer/viewer of a cloud service and not under control of the cloud provider at all. I saw this clearly when I was working at NetSuite, as we struggled to provide good response times across the entire United States. We were able to do so by carefully selecting our data center and bandwidth providers, and working with them. Is this done by all Cloud vendors? I don't know. For enterprises considering replacing internal computing with Cloud and already having well-designed internal networks, latencies will almost always be less accessing internal resources than a public cloud because the network paths are finite in number and someone has been tasked to resolve the associated latencies.
However, there are other performance factors beyond latency. To achieve that bottom-dollar price, most public clouds are built on commodity servers and networking equipment that may be adequate for a majority of applications, but are wholly incapable of addressing the needs of applications with very high I/O, memory, or CPU requirements. For customers coming to ENKI from other clouds, we've seen I/O wait percentages of 80% or more of subscribed CPU time for some applications, particularly those with intense database loads or interconnected instances. In other words, 80% of the cloud subscribers' paid resources went unused due to I/O delays! Without knowing the architectural limitations of proprietary public cloud architectures such as Amazon, I surmise that these delays could result from congestion at the shared physical server's network interface, or from storage performance requirements that were not met by the cloud provider's storage offering - among other causes. In general, as the speed of processors has outstripped that of storage and I/O, these problems have gotten worse. But the use model of Cloud, which fully occupies the hardware through virtualization and shared usage, means that they can become critical. Solving some of these problems so that Cloud users will be able to take advantage of today's faster processors will be expensive since they will require radically different hardware infrastructure than current cloud providers offer, which I expect will result in significantly higher prices for high performance Cloud Computing.
In the meantime, working with a vendor that can assist you in adapting your application to the cloud can reduce the impact of any potential performance bottlenecks by taking advantage of the underlying systems' strengths.
I don't think these limitations are a blot on Cloud Computing itself (which I'm of course excited about!), but rather a call to further discussion, development of higher-performance cloud infrastructure, and of course a frank discussion by cloud vendors of the limitations of their solutions. This will avoid customer disappointment and frustration as well allowing cloud vendors to focus on their customer's needs. Avoiding this discussion concerns me that Cloud will be labeled as an over-hyped product.
Here at ENKI we've seen the need for a new hardware architecture that can provide radically higher performance to our customers, and we're in the final stages of qualifying it. It will provide radically enhanced I/O bandwidth by speeding connections between our servers and to storage, as well as offering greatly enhanced storage speeds. This will allow our customers to deploy applications to our Cloud that formerly would have required a custom-designed datacenter and the associated costs and skills.
TrackBack URI for this entry
