NSDI 2012 Day 1

NSDI 2012@San Jose Fairmont.

Other info from here:-

Session 1 - Big Data.

Corfu - clusers of raw flash units
(also island in greece near Paxos)....
produces orderd list of events...

Network attached Flash - aves power&latency
Then do log-structured stuff on it....
perf with emulated unit does 1M event per sec....v. impressive

comparison - paxos is i/o bound...corfu isn't...

one good question about what happens if Flash guys actually deliver on some
promises - how would that change corfu - basically, not a lot.

nice work - a lot of effort

Resilient Distributed Datasets (RDDs):

A Fault-Tolerant Abstraction for In-Memory Cluster Computing
mapreduce paper:- replacement for HDFS
RDDs are really just another level of indirection with a query API
to the aggregator function -

implemented in Spark in Scala + dryadlinq like api (open sourced)

eval on example of log mining
full text search on wikipedia (50Gig) in seconds.,,

Seemed rather obvious to me - language looks quite nice though..
{see Nextar@OSDI10 - does cite - also cites CIEL:-)

Brave presenter did a live demo...which worked...

Question - what happens when not enough ram - thrash to disk...
speaker had backup slide that showed performance hit...fixing this
maybe distributing memory load/work was future work...

Camdoop: Exploiting In-network Aggregation for Big Data Applications- Paolo Costa

More mapreduce throughput speedup work
This looks at how data reduction maps to actual network topology (viz fattree etc)

Observe (as others have) that final results in many typcal apps (taken from facebook examples)
are 5% of intermediate data

So use this to do some aggregation in an intermediate stage - now try and do this
in a distributed way - problem is how to map logical disributed aggregation
onto physical topology - creates hot spots in the aggregation tree....

mismatch between logical&physical topology..
Camdoop targets reducing shuffle time by packet aggregation
takes advantage of CamCube topology
So map tree to camcube -see cute diagrams in paper

eval - built a 3^3 camcube with quad core servers + have same code in simulator with 8^3 eval
look at worst case perf (good for them) in bisection bandwidth - gain is good exactly where they want it -
show all-to-one and all-to-all - looks like about a 5 fold perf gain in common case (fb workload)

Wireless Session

1. WiFi-NC : WiFi Over Narrow Channels from MSR India

Related idea to Bogdan's thesis
(nice to see yet another talk not doing the powerpoint thing too:)
basically doing clever stuff with subcarriers/narrow band to get effiecient coding/use of channels -
this works within ISM and OFDM, but also for Whitespace stuff, with fragmented spectrum. Essentially
get more parallelism with les interference due to independence of lots of narrow channels than 1 wideband channel (kind of obvious, but they've built it, and the applicability/flexibility to also work in whitespace with fragmented spectrum is sweet)...tradeoff really is about guardbands and multipath.

Question about 1 pb is preabmble dilation. yes, is an issue

2. Catching Whales and Minnows Using WiFiNet: Deconstructing Non-WiFi Interference Using WiFi Hardware

This is about capturing what kind of non-WiFi source is causing us (a WiFi transceiver) grief.
There are some devices/sources that are minor problems (not really disruptive) and others which are a big problem (hence, repsectively, minnows & whales) - is extension of NSDI11 PIE paper (passive interference estimation)

runs airshark on multiple APs, which can pull samples. These are reported to a combiner which then identifies sources
analogue phone, zigbee or bluetooth...this system (WiFiNet Controller) uses the multiple sources of idenified interferrers and infers how many of each type there may be.

Also uses a model-based mechanism to find the probable localaton of the devices

Evalation is empirical - the errors in locationation is median 2-4 meters (worst case was 15)
Lots of interference estimation work in the paper - looks useful and solid piece of nice engineering.

[Actually, as Dr Hand pointed out, after a while the slides in this session
became somewhat overwealming:)

Content Networking - Session 3

RPT: Re-architecting Loss Protection for Content-Aware Networks - from CMU

This is the idea of repeated transmission, with de-duping to suppress redundency
[Note to self - PGM did this for reliable multicast - retranmsissions were filtered
in subtrees that had received early copies of same data packet - RPT is slightly
more expensive in the sense that you don't just keep a sequence number, but a cache of all the actual recent
data - PGM also had a mode like that, since retransmits (or first transmits for new receivers) could be from local subtree cache only:)

Eval is through measurement of Click implementation. Workload is video, and comparison is FEC v. RPT
The implemenation is (surely) very like what you need for CCN/ICN/NDN (i.e. blackadder:)

I personally think this is a mad idea:)

Serval: An End-Host Stack for Service-Centric Networking - Erik/Princeton

Now this is nice - its a shim that makes things like mobility of ends and the like much easier -
well thought through, and architected...
So if you do migration, serval's abstraction hides this nicely..

Essentially, service name+id provides level of indirection - uses late binding.
Host s/w is a classic API/Shim.
The lookup/mapping looks quite a lot like _signpost_...
Code is in kernel and zero copy, so overheads are very low - so their
performance is very good (140,000 connection setups per second, throuhgput and latency overhead is within noise!)
LOC for delta is (for example in Firefox) 70 lines of code...
Aside on SDN to the edge of the net!
SDN works on layer 2/3 abstraction - Serval generalises this, Erik argues.

A _good_ thing. Dr Hand points out its a lot like the middleware bits in DCE/Corba/DCAN etc, which is fair.

Reliable Client Accounting for P2P-Infrastructure Hybrids - MPI+Akamai

This is the next phase of evolution for Akamai, so is v. interesting - they are building out hybrids of CDN-classic + p2p - according to bruce maggs (personal communication) they have 35000 peers to assist the core CDN setup.
[one nice result is that one can do torrent-style optimally load balanced delivery within a single AS/ISP, but then use configuration of the CDN-classic system to respect the border peering/traffic agreements _across ASs/ISPs.

This paper is maily about verifying the peers don't freeload (etc).
They make use, essentially of MPI's PeerReview (SOSP 07). That sort of makes it a fairly simple trick

A NetSession system contains rules that are what are checked by the peerreview scheme

Eval:- Client & CDN's perspective - net and CPU cost and log storage - all reasonable (as you'd expect from PeerReview)

Network Robustness - Session 4

Header Space Analysis: Static Checking for Networks, stanford, UCSD

This is a formal methodish paper which then is used a bit predicate-routinglyishly
to determine what should be where and when. Seems really quite useful and practical. The use of transfer functions and the header space transforms they perform is applied over the topology to compute what can do what.

The paper was evaluated through use cases and shown to be useful:)

A NICE Way to Test OpenFlow Applications, EPFL/Princeton

This is from Marco et al (was here for a bit) and is quite nice - maybe we should look at how we'd do this similarly with
Haris' Mirage OpenFlow? Quite a lot of this work is about state space reduction to make the checking work tractable.

Toward Predictable Performance in Software Packet-Processing Platforms, EPFL

Reminds me of Active Nets work (e.g. safetynet work at Sussex and JMS' work on ML and active packets) where you bound behavioru by limiting computational expressiveness....

Comments (0) Trackbacks (0)

No comments yet.

Leave a comment

No trackbacks yet.