5. Cost-Performance Comparisons with Other Access Methods（4）
Now 2.3 million pages is the minimum number on which the 9.2 GBytes of entries of a Continuum Structure can sit, and given 1000 inserts per second, each page of such a Structure is accessed for a new insert about once every 2,300 seconds; by the Five Minute Rule it is uneconomical to keep all these pages buffered. If we consider larger nodes to hold the entries as in the Bounded Disorder file [16], this provides no advantage, for although there is a greater frequency of reference, the cost of memory to buffer the node is also greater and the two effects cancel. In general, then, a page is read into memory buffer for an entry insert and must later be dropped from buffer to make room for other pages. In transactional systems that update disk pages in place before dropping them from buffer, this update requires a second I/O for each index insert. Thus we are able to state that a Continuum Structure that does not defer updates will require at least two I/Os for each index insert, approximately the same as a B-tree.
现在230万页是一个9.2 g字节的连续体结构条目可以容纳的最小数目，并且给定每秒1000个插入，这样一个结构的每一页大约每2300秒被访问一次新插入;根据五分钟规则，保持所有这些页面的缓冲是不经济的。如果我们像在有界无序文件[16]中那样考虑使用更大的节点来保存条目，这没有任何好处，因为尽管引用的频率更高，但缓冲区节点的内存成本也更大，这两种影响会抵消。通常，一个页被读入内存缓冲区以进行条目插入，之后必须从缓冲区中删除，以便为其他页腾出空间。在事务性系统中，在将磁盘页从缓冲区删除之前更新磁盘页，这个更新需要对每个索引插入进行第二个I/O。因此，我们可以断言，一个不延迟更新的连续体结构，对于每个索引插入至少需要两个I/ o，与b -树大致相同。（有道翻译）
todo：自己翻译 仔细看一遍
