We consider key factor to manage cache servers is a combination use of content distribution and (2) duplication of popular contents.
Content distribution increase effective cache capacity by storing different contents in servers, while content duplication could increase hit rate of cache servers by storing popular contents.
I’ll explain the details from the next slide.
At the first step, it is important to increase the effective cache capacity to reduce the traffic.
So we group (cache servers and contents) with color tags. Each color tag has a specific color. These cache servers store contents only when the color of content matches the server color.
Accordingly, each cache server stores different contents explicitly, and we could increase cache capacity up to 4 times.
However, a single use of this scheme sometimes increase traffic, especially when accessing to popular contents.
When this content with blue-color tag is a popular content, many users will request it. So this content will flow the network many times, generating additional traffic.
So we duplicate popular contents by applying multiple colors to them to increase hit rates of cache servers and eliminate traffic among cache servers.
It efficiently reduce traffic among cache servers. And also it can maintain low average hops that contribute the traffic reduction.
This slide shows an example of cache distribution.
There’re 4 cache servers with color tags, and the origin server has five contents. The yellow content is popular one, so it has tags with all colors. The rest of contents have tags with a single color.
As a result, these contents with a single tag are distributed, While the popular one is duplicated everywhere. So users could achieve popular contents from their nearest cache servers, and the rest of contents from nearby servers.
It is also important to distribute server colors to reduce average hops from users. Each cache server is preliminarily colorized with a specific color like the four-color theorem.
For a case study, we colorized the cache servers by preferring longer distances between the same colors.
Moreover, we implement a color tag as a set of bits. Each bit stands for a specific color like this figure.
For example, this color tag is described with all 1-bit, because the tag has all colors. And this color tag with 3 colors is described with 1-1-1-0, including a single 0-bit. Also, when we use 4 colors, there’re 2 to the 4th power, that is 16 types of color tags.
In order to duplicate popular contents in the network, we set tags with many 1-bits to popular contents, while less popular contents will be applied tags with many 0-bits. Accordingly, we increase hit rates of cache servers.
In addition, to follow rapid access changes, We adopt a hybrid caching scheme with colored LFU area and no-color modified LRU area. Modified LRU is a kind of LRU algorithm, which achieves better hit rate than LRU.
In our colored hybrid cache, the colored LFU area stores contents with matching color tags, while the LRU area stores any contents regardless of tags’ colors.
This LRU behavior could follow rapid changes in access patterns, even when a content with a different color rapidly rises its rank.
We also evaluated the hit rate of proposed hybrid caching scheme when access pattern changes. We inserted 5 contents with no color at a time and compared the hit rate with single colored cache.
This graph shows the cache hit rate along time. When we insert new contents, the green graph, which is for single colored cache, falls off its hit rate by 13.9 percent. On the other hand, the purple graph, which is for the hybrid cache, maintain the hit rate. Our colored hybrid cache limited the degradation to only 2.3 percent.
This is because the colored cache area couldn’t store inserted contents with no matching color, while the LRU area could store them regardless of the color tags.
Thus our colored hybrid caching scheme could follow rapid change in access pattern.
2017/08/27 July Tech Festa 2017 1
博士後期課程3年 中島 拓真
中島 拓真（なかじま たくま）
2017/08/27 July Tech Festa 2017 2