AUTHORS: STIJN POLFLIET ET. AL. BY: ALI NIKRAVESH Studying Hardware and Software Trade-Offs for a Real-Life Web 2.0 Workload

Agenda Introduction - What is the research question? What is authors’ approach to address the problem? Results 06/10/2015 2/15

Introduction Online Social Networking  Facebook  Twitter  LinkedIn  Netlog Large number of users  designing the servers and data centres to support social networking is challenging Large scale data centres  very much cost driven 06/10/2015 3/15

Introduction (contd.) Factors that affect the cost of data centre:  Hardware infrastructure  Power and cooling infrastructure  Operating expenditure  Real estate Increase performance  increase cost  Reduce cost  reduce performance  RESULT  optimize performance per dollar 06/10/2015 4/15

Introduction (contd.) So, which server should be bought?  Depends on workload (interactive, batch, memory intensive,…) Idea: different workloads run on different hardware  Correct hardware for correct task 06/10/2015 5/15

Introduction (contd.) Research Question: Scope: Social networking Can we come up with a way of guiding service operators and owners of data centre to what hardware to purchase for a given workload? 06/10/2015 6/15

Analysis Social networking workload  Multiple services run on multiple servers in a distributed way in a data center Fundamental difficulty  performance of the ensemble can only be measured by modeling and evaluating the ensemble  Authors’ approach: A case study to evaluate how hardware choices affect end-user experience 06/10/2015 7/15

Case study Netlog  Is a social network available in 40 languages Architecture  WS 54%  MC 16%  DB 30% 06/10/2015 8/15

Experimental Setup 10 dual AMD Opteron 6168 servers 12 cores per CPU 64 GB of main memory HDD & SSD 06/10/2015 9/15

Web Server Results Three CPU frequency  1.9 GHz  1.3 GHz  800 Mhz Frequency has a significant impact on response time. 06/10/ / 15

Web Server Results (contd.) Number of cores per node  Four-socket system is typically more than twice as expensive as a two-socket system.  As long as the total number of cores is constant, CPU node is not affected much by node and core count 06/10/ / 15

Database Server Database server: HDD versus SDD Although this is a significant reduction in the longest response times observed, it may not justify the significantly higher cost of SSD versus HDD 06/10/ / 15

Memcached Server Average CPU load for memcached server is typically below 5% 06/10/ / 15

Usecase Hardware Purchasing  Hardware vendor suggestion:  Web server: Intel Xeon X3480, 8 GB RAM, typical HDD ($1795)  Memcached server: Intel Xeon X3480, 16GB RAM, typical HDD ($2015)  Database server: Intel Xeon X3480, 16GB RAM, SSD ($2915)  Total: $18615  Suggestion  Web server the same as above. Memcached the same as Web server but lower CPU frequency. DB the same as memcached but not SSD (Total: $15015 – 18.9% reduction)  Performance Evaluation:  50% of all requests will not experience any extra latency  For other 50%, increase from 11% to 39% 06/10/ / 15

Thanks Questions? 06/10/ / 15

Les important diagrams 1 06/10/ Comparing simulation response with actual response packets

Les important diagrams 2 06/10/ Sampling in time Traffic classified by its type

Les important diagrams 3 06/10/ Warm-up

Les important diagrams 4 06/10/ CPU load as a function of clock frequency CPU load as a func of cores