Ming YIN
Crazy Monster!
Data Center TCP
or DCTCP [1]
Mohammad Alizadehzy, Albert Greenbergy, David A. Maltzy, Jitendra Padhyey, Parveen Pately, Balaji Prabhakarz, Sudipta Senguptay, Murari Sridharany
Microsoft Research & Stanford University
SIGCOMM’10, August 30–September 3, 2010, New Delhi, India.
Ming YIN
21 - 11 - 2013
Data Center
A Hot Topic
Applications
Characteristics
TCP - Transmission Control Protocol
A Old Topic
Internet
Impairments
DCTCP
A New Topic
For Internet Applications
To solve Impairments of TCP
because of the Characteristics of Data Center
Contributions
Extract the application patterns (characteristics) of Data Center.
Find the performance impairments of TCP used in Data Center.
- Incast
- Queue Buildup
- Buffer Pressure
Propose and Implement DCTCP to solve these impairments.
- High Burst Tolerance
- Low Latency
- High Throughput
Data Center
What is a Data Center?
"A data center is a facility used to house computer systems and associated components, such as telecommunications and storage systems."
- Wikipedia [2]
Many companies are using Data Centers
LINK [4]
Why?
High Availability
Infrastructure Scalability
Cost of Connection
Security
more
Infrastructure Scalability
Cost of Connection
Security
Data Center Architecture
Cluster - Rack - Server
Commodity switches
Communications in Data Centers
Storage Query
Short flows
Big Data Computation
Short flows & Long flows
Data Update
Long flows
Data Center Workflow Characterization
A common application structure for Soft real-time applications:
Partition/Aggregate Pattern - Latency Sensitive - Bursty
Figure 1: A sample partition/aggregate pattern
Query Traffic
Partition/Aggregate Pattern - Delay Sensitive
A High-level aggregator
partitions queries to a large number of
Mid-level aggregators
partition query over
Servers
Background Traffic
Update Flows (Large flows) & Short Messages (Small flows)
Large flows: throughput-sensitive
Small flows: delay-sensitive
Flow Concurrency
Large flows, short flows and bursty query traffic
co-exist in a data center network.
Figure 2: Distribution of number of concurrent connections. [1]
Performance Impairments
Some backgrounds:
-
TCP rules Data Center traffics, 99.91% [1].
-
Shared memory switches (shallow or deep buffered).
-
Packets are queued for an outgoing interface.
-
A packet is dropped if the buffer of a port is full.
Performance Impairments of TCP
Figure 3: Three main impairments.
(a) Incast (2) Queue buildup (3) Buffer pressure
Queue - TCP
- Packets may be dropped when queue is Full
- Packets are delayed when queue is built up.
- The buffer size of port is limited since long flows take up shared memory.
- TCP is greedy in traffic consumption
- TCP reacts to congestion too late
- Explicit Congestion Notification
- TCP drop the window in half when congestion occurs
DCTCP
Goals
- Small queue occupancies
- High throughput
Approaches
- Congestion Experienced (CE) codepoint marking
- TCP congestion window size limitation
- Multi-bit feedback to Single-bit sequence of marks
- Reuse the ECN machinery
Simple Marking at the Switch
A single threshold parameter: K
ECN-Echo at the Receiver
Receiver sets ECN-Echo flag for
every ACK packets when CE is marked.
Controller at the Sender
An estimation of the fraction of packets that are marked: 𝜶
F is the fraction of packets marked in the last window of data.
g is the weight given to new samples
𝜶 ← ( 1 - g ) × 𝜶 + ( g × F )
Controller at the Sender (cont)
𝜶 means the probability that the queue size is greater than K.
Congestion window size: cwnd
cwnd ← cwnd × ( 1 - 𝜶 / 2 )
DCTCP vs. TCP
React in proportion to the extent of congestion.
Not its presence.
Table 1: How congestion window is updated with different ECN marks. [4]
Benefits
1. Incast
- Large buffer headroom → Burst fits
- Aggressive marking → Earlier reaction to the congestion
2. Queue buildup
- Small buffer occupancies → Smaller queueing delay
3. Buffer pressure
- A port's queue does not grow exceedingly large
Results
Figure 4: The Queue size of TCP and DCTCP in action [1].
Setup: Win 7, Broadcom 1Gbps Switch
Scenario: 2 long-lived flows, K = 30KB
Cluster Benchmark Traffic
Figure 4: The measured result of DCTCP and TCP in
an emulated traffic within 1 Rack of Bing cluster [5].
Implications and recent work
A fluid analysis model of DCTCP
Stability, convergence and fairness [6]
HULL (High-bandwidth Ultra-Low Latency) architecture
Head room for latency sensitive traffic [7]
Conclusion
DCTCP is proven to handle
- High Burst Tolerance
- Low Latency
- High Throughput
Features
- Small changes to TCP → Easy to understand and test
- Reuse existing mechanisms → Applicable
References
[1] Alizadeh, Mohammad, et al. "Data center tcp (dctcp)." ACM SIGCOMM Computer Communication Review 40.4 (2010): 63-74.
[2] "Data center." Wikipedia, The Free Encyclopedia. Wikimedia Foundation, Inc. 22 July 2004. Web. 10 Aug. 2004.
[3] "Map of Uptime Institute Tier Certified Data Centers." Uptime Institute, the Data Center Authority. 15 April 2011. Retrieved from http://blog.uptimeinstitute.com/2011/04/map-of-uptime-institute-tier-certified-data-centers.
[4] "Companies." Data Center Knowledge.Retrieved from http://www.datacenterknowledge.com/archives/category/companies/
[5] Mohammad Alizadeh, "Data Center TCP (DCTCP) ." IETF Talk. Retrieved from http://www.ietf.org/proceedings/80/slides/iccrg-3.pdf
[6] Alizadeh, Mohammad, Adel Javanmard, and Balaji Prabhakar. "Analysis of DCTCP: stability, convergence, and fairness." Proceedings of the ACM SIGMETRICS joint international conference on Measurement and modeling of computer systems. ACM, 2011.
[7] Alizadeh, Mohammad, et al. "Less is more: Trading a little bandwidth for ultra-low latency in the data center." Proc. of NSDI. 2012.
Questions?
Please leave your comments to this slide:
Data Center TCP
By Ming YIN
