Computer
scientists have developed an inexpensive solution for diagnosing delays
in data center networks as short as tens of millionths of
seconds—delays that can lead to multimillion-dollar losses for
investment banks running automatic stock trading systems. Similar
delays can delay parallel processing in high performance cluster
computing applications run by Fortune 500 companies and universities.
|
Computer scientists have developed an inexpensive solution for diagnosing networking delays in data center networks as short as tens of millionths of seconds—delays that can lead to multimillion-dollar losses for investment banks running automatic stock trading systems.
|
University
of California, San Diego and Purdue University computer scientists
presented this work on August 20, 2009 at SIGCOMM, the premier
networking conference.
The new approach offers the
possibility of diagnosing fine-grained delays—down to tens to
microseconds—and packet loss as infrequent as one in a million at every
router within a data center network. (One microsecond is one millionth
of a second.) The solution could be implemented in today’s router
designs with almost zero cost in terms of router hardware and with no
performance penalty. The UC San Diego and Purdue University computer
scientists call their invention the Lossy Difference Aggregator.
“A
lightweight network monitoring approach such as ours allows you to
pinpoint the source of the performance degradation and identify the
problem routers,” explained SIGCOMM 2009 paper author Kirill Levchenko,
a UC San Diego post-doctoral researcher who recently earned his Ph.D.
in computer science at UC San Diego.
“This is
stuff the big traders will be interested in,” said George Varghese, a
computer science professor at the UC San Diego Jacobs School of
Engineering and an author on the SIGCOMM paper, “but more importantly,
the router vendors for whom such trading markets are an important
vertical.”
If an investment bank’s algorithmic stock
trading program reacts to information on cheap stocks from an incoming
market data feed just 100 microseconds earlier than the competition, it
can buy millions of shares and bid up the price of the stock before its
competitors’ programs can react, the computer scientists say.
While
the network links between Wall Street and investment banks’ data
centers are short, optimized and well monitored, the performance of the
routers within the data centers that run automated stock trading
systems are difficult and expensive to monitor. Delays in these
routers, also known as latencies, can add 100s of microseconds,
potentially leading to millions of dollars in lost opportunities.
“Every
investment banking firm knows the importance of microsecond network
delays. Because routers today aren’t capable of tracking delays through
them at microsecond time scales, exchanges such as the London Stock
Exchange use specially crafted external boxes to track delays at
various key points in the data center network,” said Alex Snoeren, a
computer science professor at the UC San Diego Jacobs School of
Engineering and an author on the SIGCOMM paper.
But these
external systems are generally too large and expensive to be added to
every router in a data center network running an automated stock
trading system. This makes it difficult for the network managers to
identify and locate problematic routers before they cost the company
large amounts of money, the computer scientists say.
“Our
hope is that this approach will allow router vendors to add fine scale
delay and loss tracking, at almost zero cost to router performance,
perhaps obviating the desire for expensive external network monitoring
boxes at every router,” said Ramana Kompella, the first author on the
SIGCOMM paper and a computer science professor at Purdue University.
Kompella earned his Ph.D. in computer science at UC San Diego in 2007.
The
SIGCOMM 2009 paper presents simulations and proof-of-concept code for
measuring latencies down to tens of microseconds and losses that occur
once every million packets.
|
“A lightweight network monitoring approach such as ours allows you to pinpoint the source of the performance degradation and identify the problem routers,” explained SIGCOMM 2009 paper author Kirill Levchenko, a UC San Diego post-doctoral researcher who recently earned his Ph.D. in computer science at UC San Diego.
|
“The next step would be to
build the hardware implementation, we are looking into that,” said
Kompella, who plans to continue pioneering research in fault diagnosis
at Purdue.
This work highlights a fundamental shift
happening across the Internet. As computer programs—rather than
humans—increasingly respond to streams of information moving across
computer networks in real time, millionths of seconds matter.
Algorithmic stock trading systems are just one example. Extra
microseconds of delay can also mean slower response times across
clustered-computing platforms, which can slow down
computation-intensive research, such as drug discovery projects.
“When
it comes to fault isolation, networks are a big black box. You put
packets in on one side and you get them out the other side,” explained
SIGCOMM paper author Kirill Levchenko, a UC San Diego post-doctoral
researcher who recently earned his Ph.D. in computer science at UC San
Diego. “A lightweight network monitoring approach such as ours allows
you to pinpoint the source of the performance degradation and identify
the problem routers.”
Lossy Difference Aggregator
Simple counters and clever thinking are at the heart of the Lossy Difference Aggregator.
The
classical way to measure latency is to track when a packet arrives at
and leaves a router, take the difference of these times, and average
over all packets that arrive over a fixed time period, such as one
second. However, a typical router may process 50 million packets in a
second, and keeping track of each packet’s arrival and departure is a
daunting piece of bookkeeping. It may seem that a simple approach is to
sum all the arrival times in one counter, sum all the departure times
in another counter, subtract the two counters and divide by number of
packets. Unfortunately, this simple “aggregation” idea fails when a
packet is lost within a router (which commonly happens). In that case,
the lost packet arrival time is included but its departure time is not,
throwing the whole estimate wildly out of whack.
Instead of
summing the arrival and departure times of all packets traveling
through a router, the computer scientists’ system randomly splits
incoming packets into groups and then adds up arrival and departure
times of each of the groups separately. As long as the number of losses
is smaller than the number of groups, at least one group will give a
good estimate.
Subtracting these two sums (from the groups
that have no loss) and dividing by the number of messages provides an
estimate of the average delay with very little overhead—just a series
of lightweight counters.
With this invention built into
every router, a data center manager should be able to quickly pinpoint
the offending router and interface that is adding extra microseconds of
delay or losing even a few packets in a million, explained Levchenko.
“This
is diagnostic tool, a potentially extremely important one. You don’t
want to just know that you have a network problem, you want to know
which router and which application is causing the problem,” said
Snoeren.
The network manager can then upgrade the router
or link, or reassign an offending application that is sending message
bursts to another processing path.
By contrast, today’s
routers can be made to log messages; but looking through logs of
millions of messages to pinpoint delay problems is like looking for a
needle in a haystack.
“If implemented, this kind of
approach should enable investment bankers to turn their attention to
tuning their algorithmic trading programs to make more intelligent
investments, instead of worrying about delays through obscure routers,”
said Varghese.
Original article
