By the Numbers: AWS' Latency Vulnerability
When cloud-based application performance depends on latency, going with the number two cloud service provider might be the best option. Azure consistently bests its most prominent rivals when it comes to network delay.
In this installment of By the Numbers, we focus on latency in three of the leading cloud networks in North America: Amazon Web Services Inc. (AWS), Azure Solutions Inc. (Microsoft) and SoftLayer Technologies Inc. (IBM). To keep things simple, we narrowed our comparison to performance measurements across a few of the larger service providers (SPs). These include AT&T Inc. (NYSE: T), CenturyLink Inc. (NYSE: CTL), Charter Communications Inc. , Comcast Corp. (Nasdaq: CMCSA, CMCSK), Frontier Communications Corp. (NYSE: FTR), MCI LLC (Verizon), and Time Warner Cable Inc. (NYSE: TWC). (See By the Numbers: Network Latency in Texas.)
A voluminous amount of data -- collected and presented by Cedexis -- is available. Our intent is to keep the comparisons manageable, but that also means comparisons are not always one-to-one. For example, most service providers have single regional autonomous system numbers (ASNs), but Time Warner Cable runs at least six separate ones; we chose to look at only the ones with the most traffic of the lot. (Charter, which now owns TWC, has made no public statements about integrating TWC's ASNs.)
(View larger version)
The graph shows that when it comes to latency from these eyeball networks, Azure is almost always better than AWS. For that matter, in many cases, so is Softlayer -- for this 30-day period, at least. Traffic patterns are dynamic, and cloud providers upgrade their networks continuously.
Latency is hardly the only measure of a cloud network, but it is among the more critical. Excessive latency can degrade streaming video. When loading web pages, excessive latency slows page load time (PLT). It is not uncommon for the web pages of some news sites to be comprised of hundreds of elements; an extra delay of 50 ms -- or even 25 ms -- compounded over hundreds of elements can lead to a page loading noticeably more slowly.
Not every application requires the lowest latency available. AWS has a formidable lead in the cloud market, and has developed a richer palette of services that many customers no doubt rely upon. Nonetheless, the statistics indicate that AWS has a vulnerability, and that's in applications where speed is of the essence.
Generally speaking, accessing the cloud is significantly faster in the East than in the West of North America. In the East, the average latency is 55.67 milliseconds (all measurements are at the 50th percentile); in the West, the average latency is 89.61 ms.
Yes, there are some caveats to consider when examining these averages, but the distortions are minor. For instance, we measured the latency when CenturyLink connects with the cloud in the East -- 99 ms -- performance that drags down the average for connectivity in the East. But given that CenturyLink's operations are largely confined to the northwest, it can be expected to experience greater latencies when accessing clouds physically located in the East. Factoring in those connections inflates the average latency in the East, but taking that into account actually reinforces the superior performance in East clouds.
Cloud access through MCI's network (a.k.a. Verizon) in the East is notably exceptional. Latency is in the low 20s regardless of the cloud (21 ms with Softlayer; 23 ms with both AWS and Azure). No other cloud/SP combination comes anywhere near latencies that low, and that includes MCI itself with its cloud partners in the West.
Conversely, AT&T in the East clearly lags its competitors (pun intended).
Taking a look at West Clouds in the chart above, everyone's performance is close to the mean, with the exception of outlier TWC. To be fair, this TWC ASN performs well in the East; in the West it is almost certainly experiencing the same inter-regional lag that makes Western-based CenturyLink's numbers in the East look so much worse.
Cedexis measures every cloud instance in the world billions of times a day from every ISP and cable network. About 7 billion measurements a day are taken of latency, throughput and availability to each of these clouds using Real User Measurements (RUM).
— Brian Santo, Senior Editor, Components, T&M, Light Reading