|
 코드번호 : 4 |
|
| 발표자 : 한동수 |
|
| 소속 : KAIST |
|
| 부서 : 전기 및 전자공학과 |
|
| 직위 : 교수 |
|
| 세션시간 : 2015년 6월 23일 13:40-15:40 |
|
| 발표자약력 : |
2012.12 Carnegie Mellon University (박사)
2004.2 KAIST 전자 전산학과 졸업 (학사)
2013.1 ~ 2013.5 박사후 연구원 Carnegie Mellon University
2013.6 ~ 현재. KAIST 전기 및 전자공학과 조교수 |
|
|
| 강연요약 : |
Cloud computing has been tremendously successful in providing the scalability needed for popular Internet-based services such as
Facebook, Google, and Netflix. Popular Internet-based services that we use everyday rely on a public or private cloud platform to provide services world-wide. To scale out the system performance, these services utilize hundreds of thousands to millions of servers that cost billions of dollars. Effectively utilizing individual resources in the cloud becomes crucial in reducing the cost (and the energy consumption) at this scale.
In this talk, we look at how to improve the performance of a single machine within the cloud. One of the key problems that limits the performance is that applications and networking stacks are not design to scale well in multicore environments. Modern machines have tens of cores and multiple 10Gbps Ethernet links. However, existing designs cannot effectively utilize these resources. In this talk, we focus on improving the performance of two essential building blocks, an in-memory key-value store and the TCP stack, that Internet-based services commonly rely on. By employing multi-core aware designs and by-passing the OS kernel to eliminate its overhead, we show that we can dramatically increase their performances by up to 13.5x and 3x respectively. |
|
|
|
