Check out the new USENIX Web site. next up previous
Next: Application Benefits Up: Evaluation / Live Traffic Previous: Reliability and Availability

Overhead Analysis


To analyze CoDNS's overhead, we examine the remote query traffic generated by the CoDeeN live activity. For this workload, CoDNS issued 11% to 85% of the total lookups as remote queries, as shown in Figure 19. The variation reflects the health of the local nameserver, and less stable nameservers require more remote queries from CoDNS. Of the six nodes that had more than 50% remote queries, all experienced complete nameserver failure at some point, during which remote queries increased to over 100% of the local requests. These periods skew the average overhead.

We believe that the additional burden on nodes with working DNS is tolerable, due to the combination of our locality-conscious redirection and already high local nameserver hit rates. Using our observed median overhead of 25% and a local hit rate of 80% - 87% [9], the local DNS will incur only 3.25 - 5.00% extra outbound queries. Since remote queries are redirected only to lightly loaded nodes, we believe the extra lookups will be tolerable on the peer node's local nameserver.

We also note that many remote queries are not answered, with Figure 19 showing this number varies from 6% to 31%. These can be due to WAN packet losses, unresolvable names, and remote node rate-limiting. CoDNS nodes drop remote requests if too many are queued, which prevents a possible denial of service attack. CoDNS peers never reply if the request is unresolvable, since their own local DNS may be failing, and some other peer may be able to resolve the name.

The queries in which CoDNS ``wins'', by beating the local DNS, constitute 2% to 57% of the total requests. On average, 9% of the original queries were answered by the remote responses, removing 47% of the slow response portion in the total lookup time shown in the Figure 15(b). Of the winning remote responses, more than 80% were answered by contacting the first peer, specified as ``win-by-1'' in Figure 20. Of all winning responses, 95% are resolved by the first or second peer, and only a small number require contacting three or more peers. This information can be used to further reduce CoDNS's overhead by reducing the number of peers contacted - if it has not been resolved within the first three peers, then further attempts are unlikely to resolve it, and no more peers should be contacted. We may explore this optimization in the future, but our current overheads are low enough that we have no pressing need to reduce them.

In terms of extra network traffic generated for remote queries, each query contains about 300 bytes of a request and a response. On average, each CoDNS on a CoDeeN node handles 414 to 10,287 requests per day during the week period, amounting to 243KB to 6027KB. CoDNS also consumes heartbeat messages to monitor the peers each second, which contains 32 bytes of data. In sum, each CoDNS on a CoDeeN node consumes on average 7.5 MB of extra network traffic per day, consuming only 0.2% of total CoDeeN traffic in relative terms.

Figure 19: Analysis for Remote Lookups
\epsfig{file=figs/0412_0417_remote_boxfill.eps,width=3.25in,height=1.6in}

Figure 20: Win-by-N for Remote Lookups
\epsfig{file=figs/0412_0417_good_remote_boxfill.eps,width=3.25in,height=1.6in}

next up previous
Next: Application Benefits Up: Evaluation / Live Traffic Previous: Reliability and Availability
KyoungSoo Park 2004-10-02