Sorin Faibish

Jitter-free co-processing on a prototype exascale storage stack

In the petascale era, the storage stack used by the extreme scale high performance computing community is fairly homogeneous across sites. On the compute edge of the stack, file system clients or IO forwarding services direct IO over an interconnect network to a relatively small set of IO nodes. These nodes forward the requests over a secondary storage network to a spindle-based parallel file system. Unfortunately, this architecture will become unviable in the exascale era. As the density growth of disks continues to outpace increases in their rotational speeds, disks are becoming increasingly cost-effective for capacity but decreasingly so for bandwidth. Fortunately, new storage media such as solid state devices are filling this gap; although not cost-effective for capacity, they are so for performance. This suggests that the storage stack at exascale will incorporate solid state storage between the compute nodes and the parallel file systems. There are three natural places into which to position this new storage layer: within the compute nodes, the IO nodes, or the parallel file system. In this paper, we argue that the IO nodes are the appropriate location for HPC workloads and show results from a prototype system that we have built accordingly. Running a pipeline of computational simulation and visualization, we show that our prototype system reduces total time to completion by up to 30%.

A new approach to file system cache writeback of application data

In this paper we propose a new paradigm and algorithms to address cache writeback performance in servers and storage arrays. As servers and storage processors move to multi-core architecture, with ever increasing memory caches, the cost of flushing these caches to disk has become a problem. Traditional watermark based algorithms currently used in many storage arrays and NAS servers have a problem keeping up with the higher speeds of incoming application writes, often resulting in a performance penalty. The servers cache is generally used for hiding high disk latencies associated with file system data. In general, metadata performance was optimized, while application data was considered less sensitive to high latencies and was given lower priority or was written directly to disk. The new algorithms proposed here change the application data writeback from using watermark based flush to something that approximates the rate of the incoming application I/Os. The problem is more critical for network file systems where the complex client/server protocols can make writeback a serious performance barrier, particularly in light of very large I/Os and the lack of application commits. Our proposed algorithms are applicable to local file systems and remote servers as well as to storage arrays. We show test results based on dynamic traces of real file system dirty pages in the buffer cache and prove that rate based cache writeback algorithms are the most efficient replacement for watermark based flushing.

Storage virtualization using a block-device file system

The design and organization of modern file systems has been traditionally driven by practical considerations related to the physical properties of computer disks Storage virtualization makes such considerations largely irrelevant, and file-system designs based on them perform sub-optimally in a virtual storage environment. One important example of this phenomenon is the relationship between disk seek times and the placement and organization of file system meta-data. In this paper we show that traditional approaches to organizing meta-data in file systems are closely related to assumptions about the physical properties of disks and that for this reason traditional file systems fail to materialize the full benefits of storage virtualization. We go on to propose a different file system organization of data and meta-data designed to exploit the power of virtualized storage.

BAD-check: bulk asynchronous distributed checkpointing

Leadership-scale scientific simulations running as tens of thousands of tightly-coupled MPI processes are vulnerable to interruption due to a single process or node failure. Due to the dependence of each state calculation on the successful completion of each of the prior state calculations, checkpoint-restart is the most widely-used technique to achieve fault tolerance. To write a consistent view of distributed state as a checkpoint, applications typically synchronize and pause while writing data to persistent media. In this paper we present a transactional protocol that enables asynchronous distributed creation of checkpoint data sets, and describe the conditions under which it is beneficial. With simulations, we demonstrate that scientific applications exhibiting computational variance without frequent synchronization can use our protocol to either reduce run time by up to 27% or reduce required storage system capability by up to 40%.

Seamless Audio Splicing for ISO/IEC 13818 Transport Streams

This paper addresses the processing of packetized audio elementary streams (ES) during the splicing of ISO/IEC 13818 (MPEG-2) transport streams (TS). An algorithm is developed that produces a continuous audio ES across the splicing point, avoids all noticeable audio artifacts, and hence provides a seamless audio splice without the need to decode and reencode audio ESs. Perfect continuity with respect to the presentation time stamps (PTS) and careful management of the audio buffer level at all times lead to a glitch-free audio playout for an unlimited number of splices. The proposed algorithm parses the packetized audio ES up to the audio access unit (AAU) level and is very simple to implement.

A distributed caching algorithm for streaming uncompressed HD video

This poster introduces a new distributed caching algorithm used by the installable Celerra MPFS file system client to allow streaming high throughput High-Definition video streams. Large disk latencies in high throughput streaming result in a high video frame drop rate. Dropped frames rate increases with the video throughput. The algorithm proposed in this paper uses new caching ideas inspired by signal processing techniques to address dropped frames. Our techniques remove the dropped frames effect entirely, while optimizing the aggregate throughput of the storage array.

Objective image quality metrics for dct-based video compression

In this paper, three objective image quality metrics are introduced. These metrics are designed to quantify the severity of the three major types of artifacts associated with block-DCT-based image/video compression schemes, in a manner compatible with HVS. Experimental results from some commercially available encoding applications are reported in a vendor transparent manner, to illustrate the practical usefulness of the proposed metrics to assess compressed image quality aside from their theoretical significance.