The Blog of James: Where does most latency occur? (Poll Results)
The December 2007 poll at Low-Latency.com asked where people think latency enters into the system - and while the responses were plentiful and varied - the numbers raise even more questions. See the graphic below.
23 percent of you thought “Datafeed Handlers” was the principal source of latency. This is a fundamental feature of all data processing systems - the acquisition of data. These feed handlers deal with the protocol of the communications lines and nuances of getting the data into usable form. What do you think datafeed handlers do and how fast do you think they operate? People I’ve talked to say that data acquisition - for market data and the like - can take hundreds of milliseconds. That sounds like an awful lot of time.
Middleware - that middle ground of message processing, functionality, and support services - accounts for latency in 38 percent of the systems out there? WOW! I thought that was supposed to be the “lightweight” framework to facilitate higher-level functions. This should be happening so fast it isn’t measurable - certainly not compared to data acquisition. Could it be that our drive for normalizing data, services, and processing has caused us to introduce an erstwhile undesired delay in our ability to make decisions and act on them? Is this latency in your software - or something you bought from a third-party vendor? How will this be addressed? Or is there no hope?
CEP or Complex Event Processing seemed to only cause 8 percent of you an issue with respect to latency. Apparently people have off-loaded the heavy lifting to the other main components! If the core decision making components are not introducing a signficant latency problem then one of few things has happened - 1) people have already spent a lot of money on addressing this issue (how many nodes do you have?), 2) making a decision isn’t complex once you have all the data, or 3) this is a core area where firms and institutions actually have control of the situation and can address latency through direct means.
But I find that institutions are putting more and more into their central decision-making apparatus and spending more-and-more money - all in the name of low latency. How is this justified? Doesn’t Amdahl’s law tell us that improving a portion of the system that is already fairly inconsequential (with respect to latency) will result in miminal improvement in the overall performance?
Databasing - persistently storing data - is a latency concern for 31 percent of you. And well it should be! But, is it necessary to have databasing as part of the “critical path” for latency? Certainly for transactional events where guaranteed delivery and rollback are necessary! But what about other functions? Are there not a myriad of memory-based databases, or latent-storage systems available to remove this from one’s concern? Or are transactional systems what “it” is all about?
Take a moment to comment and let me know what you think the REAL problem is.
January 15th, 2008 at 2:57 pm
I agree with you James - 38% for Middleware decerves a WOW !
I think these results are biassed by the marketing budgets of the various “low latency” middleware vendors. The middleware sector is hotting up and one key axis of competition is latency. But this result is also good news for firms (such as mine!) that sell middleware latency monitoring tools. To manage latency, you have to be able to measure it.