Understanding and Designing New Server Architectures for Emerging Warehouse-Computing Environments
From AcaWiki
Citation: Kevin Lim, Parthasarathy Ranganathan, Jichuan Chang, Chandrakant D Patel, Trevor Nigel Mudge, Steven K. Reinhardt (2008/06) Understanding and Designing New Server Architectures for Emerging Warehouse-Computing Environments. International Symposium on Computer Architecture (RSS)
DOI (original publisher): 10.1145/1394608.1382148
Semantic Scholar (metadata): 10.1145/1394608.1382148
Sci-Hub (fulltext): 10.1145/1394608.1382148
Internet Archive Scholar (search for fulltext): Understanding and Designing New Server Architectures for Emerging Warehouse-Computing Environments
Download: https://dl.acm.org/doi/abs/10.1145/1394608.1382148
Tagged: Computer Science
(RSS) computer architecture (RSS)
Summary
Placeholder
Elsewhere
Problem
- Internet companies are building huge datacenters.
- Little academic prior work analyzes those workloads, including benchmarks.
Solution
- Benchmark suite
- Websearch
- Workload characteristics: unstructured data processing
- Application: Nutch search engine, using data from Wikipedia and DMOZ, hosted on [https://en.wikipedia.org/wiki/Apache_Tomcat Apache Tomcat
- Metric: Requests per second
- Why not latency?
- Webmail
- Workload characteristics: interactive internet services
- Application: SquirrelMail talking to courier-imap and exim, based on MS Exchange Server's "heavy usage" profile, and real size data from U Mich.
- Ytube
- Workload characteristics: rich media hosting
- Application: SPECweb2005 with profile based on YouTube data.
- Metric: Requests per second
- MapReduce
- Workload characteristics: analytics requests
- Application: Word count and random writes
- Metric: Latency
- Websearch
- Models
- Quantities: perf / $ (total cost of ownership assuming 3-year deprecation cycle), perf / W, perf / (infrastructure $), perf / (power & cooling $).
- Performance model: Aggregated from individual node performance
- Cost model: Hardware cost + power & cooling cost
- . See paper for details.
- Sources: publicly available prices, private correspondance
- Questions:
- Can using low-end processor help?
- HPC uses high-end "fat" nodes; if switch to "dumb" nodes, one could have more for the same price.
- Result: Depends on the benchmark. Yes if memory- or IO-bound.
- Can changing the cooling system help?
- Yes
- Can shared memory help?
- It's too hard to give each node enough memory for the worst-case working set; instead, they should pool their resources hierarchically. Probably each node won't have the worst-case at the same time, so one node can use another's vacant memory.
- Two-level memory hierarchy with first-level sized to 25% of the baseline memory seems to work well.
- Kind of. More gains possible in the future.
- Can using low-end hard-drive help?
- Only if their prices drops in the future.
- Can using low-end processor help?
Discussion
- Further research needed to determine applicability of benchmark suite.
- More work is needed to validate performance model
- They use Amdahl's law as a simplifying assumption.