This is an automated email from the git hooks/post-receive script. It was generated because a ref change was pushed to the repository containing the project "". The branch, master has been updated via da20ccea8dc537940ba5bce59aa3b073b27fd3d9 (commit) from 0e193b0a5ff6f05c696c1ffab1d793fd0099122e (commit) Those revisions listed above that are new to this repository have not appeared on any other notification email; so we list those revisions in full, below. - Log ----------------------------------------------------------------- commit da20ccea8dc537940ba5bce59aa3b073b27fd3d9 Author: cengiz <cengiz@snslmini.(none)> Date: Sat Aug 31 06:52:11 2013 -0400 Added citations to HPC application checkpoint section ----------------------------------------------------------------------- Summary of changes: papers/asg-osd/paper.bib | 26 ++++++++++++++++++++++++++ papers/asg-osd/sections/drivers.tex | 23 +++++++++++------------ 2 files changed, 37 insertions(+), 12 deletions(-) Diff of changes: diff --git a/papers/asg-osd/paper.bib b/papers/asg-osd/paper.bib index 1425a66..c752414 100644 --- a/papers/asg-osd/paper.bib +++ b/papers/asg-osd/paper.bib @@ -732,3 +732,29 @@ title = {{Design of an Intelligent Object-based Storage device}}, address = {Piscataway, NJ, USA}, keywords = {Data bases, deadlock detection, distributed data bases, graph theory, graph theory, Data bases, deadlock detection, distributed data bases}, } + +@inproceedings{ali:scalable-iofwd, + author = {Nawab Ali and Philip Carns and Kamil Iskra and Dries Kimpe and Samuel Lang and Robert Latham and Robert Ross and Lee Ward and P. Sadayappan}, + title = {{Scalable I/O Forwarding Framework for High-Performance Computing Systems}}, + booktitle = {Proceedings of IEEE Conference on Cluster Computing, New Orleans, LA}, + month = {September}, + year = {2009} +} + +@inproceedings{Atikoglu:2012:WAL:2254756.2254766, + author = {Atikoglu, Berk and Xu, Yuehai and Frachtenberg, Eitan and Jiang, Song and Paleczny, Mike}, + title = {Workload analysis of a large-scale key-value store}, + booktitle = {Proceedings of the 12th ACM SIGMETRICS/PERFORMANCE joint international conference on Measurement and Modeling of Computer Systems}, + series = {SIGMETRICS '12}, + year = {2012}, + isbn = {978-1-4503-1097-0}, + location = {London, England, UK}, + pages = {53--64}, + numpages = {12}, + url = {http://doi.acm.org/10.1145/2254756.2254766}, + doi = {10.1145/2254756.2254766}, + acmid = {2254766}, + publisher = {ACM}, + address = {New York, NY, USA}, + keywords = {key-value store, memcached, workload analysis, workload modeling}, +} diff --git a/papers/asg-osd/sections/drivers.tex b/papers/asg-osd/sections/drivers.tex index f4a94f0..4b9a7f1 100644 --- a/papers/asg-osd/sections/drivers.tex +++ b/papers/asg-osd/sections/drivers.tex @@ -97,21 +97,20 @@ For example, say why this can't be easily implemented by T10 attribute pages} \subsection{HPC Application Checkpoint} -\cengiz{Complete citations in this section} - \ac{HPC} application workloads are characterized by bursty, highly concurrent, -write-intensive I/O patterns~\cite{TODO}. In particular, many scientific -simulations periodically write timestep data for subsequent analysis or -periodically write checkpoint data for application resilience. In these -scenarios, all application processes typically write simultaneously to the -same shared data set as shown in Figure~\ref{fig:conc-write}. Although the -application processes are coordinated and do not generally write to +write-intensive I/O patterns~\cite{Atikoglu:2012:WAL:2254756.2254766}. +In particular, many scientific simulations periodically write timestep data +for subsequent analysis or periodically write checkpoint data for application +resilience. In these scenarios, all application processes typically write +simultaneously to the same shared data set as shown in Figure~\ref{fig:conc-write}. +Although the application processes are coordinated and do not generally write to overlapping byte ranges in the file, the access patterns may be highly interleaved and are not necessarily block aligned. Optimizations such as -two-phase I/O~\cite{TODO} and I/O forwarding~\cite{TODO} can be used to -mitigate the level of concurrency observed by the storage system, but data -must still be written by many processes in order to leverage enough I/O paths -to meet bandwidth requirements. +two-phase I/O~\cite{delrosario:two-phase, thakur:ext2phase} and I/O +forwarding~\cite{ali:scalable-iofwd} can be used to mitigate the level of +concurrency observed by the storage system, but data must still be written +by many processes in order to leverage enough I/O paths to meet bandwidth +requirements. \td{Text below really doesn't make sense/isn't clear. Maybe instead (or in addition) focus on how forks can be used to get each rank a private space hooks/post-receive --