TL;DR
This paper investigates how quickly the Domain Name System (DNS) can actually react to network changes, finding that it's often too slow for critical tasks like rerouting traffic during failures. The authors propose proactive techniques to improve DNS responsiveness through cooperation between content providers and their clients.
What Problem Does This Paper Address?
Large online services and content providers increasingly rely on techniques like mirroring, content delivery networks (CDNs), and multi-homing to ensure their services are available and perform well. The Domain Name System (DNS) is a key enabler for these strategies, acting as a "late binding" mechanism to direct users to the best available server or network path. The assumption is that DNS can quickly adapt to changing network conditions, such as server failures or sudden traffic surges (flash crowds). However, this paper questions the actual responsiveness of DNS in practice, particularly when network control requires rapid adjustments.
Methods
The authors employed a two-pronged approach to measure DNS responsiveness:
- Web Event Measurements: They analyzed access logs from several large web events (sports events) hosted on a distributed infrastructure. By correlating client access logs with administrator logs detailing DNS update times, they could identify how quickly clients stopped accessing old servers after a DNS change that pointed them to new ones.
- CDN Measurements: They examined DNS request logs from servers within the Akamai CDN. This involved analyzing the behavior of local DNS servers (LDNSes) when querying authoritative DNS servers for content. They specifically looked at whether LDNSes respected the Time-To-Live (TTL) values set for DNS records, which dictate how long a record can be cached before a fresh lookup is required.
Datasets
- Web Event Logs: Cache access logs from three large sporting events in April, June, and July 2003. These logs were combined with administrator logs indicating DNS update timestamps.
- Akamai CDN Logs: DNS request logs from Akamai's high-level name servers, capturing requests from LDNSes. This was supplemented by tracking the subsequent requests made by a sample of these LDNSes to Akamai's low-level name servers.
Key Insights and Ideas
- DNS is a Coarse-Grained Tool: The core finding is that DNS, while ubiquitous, is not agile enough for applications demanding rapid responses to network events. Its inherent caching mechanisms and the behavior of many DNS resolvers limit its effectiveness for real-time control.
- TTL Violations are Significant and Long-Lasting: A substantial fraction of clients and LDNSes do not adhere to DNS TTLs. Crucially, when they do violate TTLs, they often do so for extended periods, sometimes hours or even a full day, rendering DNS control ineffective during those times.
- Not a Single Culprit: The analysis suggests that no single DNS server implementation (like BIND) is solely responsible for TTL violations. Instead, the issue appears to be more widespread, involving various implementations and network configurations.
- Cooperation is Key for Improvement: The paper advocates for a cooperative model where content providers and their major clients work together. This partnership can enable more proactive DNS control mechanisms that go beyond simple TTL settings.
Key Results
- Web Events: Between 53% and 67% of client traffic was redirected quickly after a DNS update. However, a significant portion of remaining traffic continued to access old servers for extended periods.
- Web Events (Violations): Up to 47% of clients and LDNSes collectively exhibited TTL violations. The majority of these violations lasted for more than 2 hours, with 75-85% of violating clients continuing to use stale DNS entries for over 2 hours.
- CDN (LDNSes): Approximately 14% of observed LDNSes violated TTLs on NS records. Nearly 70% of these violating LDNSes showed violations exceeding 1 hour, and about 25% exceeded 5 hours.
- NAC Analysis: A higher proportion of violating LDNSes were found in smaller network clusters (NACs), particularly those with prefix lengths of /21 to /24, suggesting potential configuration or deployment issues in these smaller networks.
Limitations and Future Work
- Limitations:
- The analysis of web events could not distinguish between violations caused by client applications and those caused by noncompliant LDNS servers.
- The measurement period for web events extended only about a day after the TTL expiry, so violations beyond that timeframe could not be captured.
- The analysis of LDNS behavior was biased towards servers making frequent requests, potentially undercounting violations from less active LDNSes.
- The BIND version analysis was limited to specific implementations and might not capture all non-BIND servers.
- Future Work:
- Conduct more active studies using probes to directly measure LDNS adherence to TTLs.
- Further investigate the root causes of noncompliance.
- Develop and evaluate the proposed architectural techniques for proactive DNS control, such as push-based invalidation and adaptive leases.
Summary updated