INDEX
| Raweb 2003 / Project-Team : reso
|
HELP INDEX |
|
Key words: execution environments, programmable and active networks.
We have proposed a new execution environment called Tamanoir, which focuses on performance problems of active and programmable network equipments and dynamic deployment of services [13] . Targeted equipments are deployed in access networks around high performance (Gbit/s) backbones. These networks must face heterogeneity problems in terms of equipments and bandwidth.
The Tamanoir architecture is designed to be a high performance active router able to be deployed around high performance backbones. This approach concerns both a strategic deployment of active network functionalities around backbone in access layer networks and providing a high performance dedicated architecture.
Tamanoir Active Nodes (TAN) provide persistent active nodes supporting various active services applied to multiple data streams at the same time. Both main transport protocols (TCP/UDP) are supported by the TAN for carrying data. We rely on the user space level of the 4 layers of the Tamanoir architecture (Programmable NIC, Kernel space, User space and Distributed resources) in order to validate and to deploy our active collaborative cache services.
The high performance Tamanoir architecture has been implemented on a cluster-based infrastructure and supports active services inside the Linux kernel and on distributed resources [31] [17] [18] .
Experimental tests have been made around high performance backbone (RNRT VTHD++ project), and for alternative support of Grid network infrastructure (RNTL e-Toile).
Our environment has been used and deployed inside various applications context like :
Active Web [33] to efficiently support on-the-fly protocol change for web sessions;
Collaborative web caches [34] to deploy intelligent and lightweight caches inside the network;
Active Logistical networks [25] to provide efficient storage functionalities inside the network for multimedia streams.
Key words: storage, active networks.
Logistical networks provide efficient distributed storage solutions inside networks. The Internet Backplane Protocol (IBP) developed by LoCI laboratory (Univ. Tennessee Knoxville, USA) allows the sharing of storage resources through wide area networks. IBP is based on data blocs (disk, memory...) and proposes a complete data depot solution. IBP depots are distributed between sites and can be accessed remotely by data streams to deploy a global storage service
We have studied the integration and merging of logistical networks inside our active network solutions in order to allow active services to efficiently store data on the fly [25] . From this new proposed architecture, we have merged and developed an active logistical equipment based on Tamanoir execution environment and IBP.
Key words: web caches, collaboration, active networks.
During the DEA internship of Sidali Guebli jointly supervised between J.M. Pierson (LIRIS, INSA Lyon) and L. Lefèvre, we studied the support of active networks to lightweight communications infrastructure for collaborative web caches. Some of the difficulty lies in the limited resources we want to deploy on the active nodes (in terms of CPU, memory and disk). But, we clearly benefit from active networks support by transparently deploying active caches through data path without modifying and re-configuring Web clients and servers. Collaborative web caches services have been developed in Java inside Tamanoir EE. These active cache services can be dynamically modified and communicates in point to point way through control communication channel between active nodes. Active dedicated services have been developed and deployed and validated on local experimental platform [34] .
Load balancing in cluster-based active network equipments
Contributed by: Pierpaolo Giacomin, Laurent Lefèvre.Key words: software router, cluster, load balancing.
As programmable network equipments allow deployment of heterogeneous services, we propose new solutions to efficiently balance equipments based on clusters. We propose new load balancing policies added to the Linux Virtual Server Project (LVS).
Key words: Reliable multicast, programmable networks, active services.
Active services for reliable multicast proposed so far in the research community consisted in the cache of data packets and the feedback aggregation. Caching data is very costly to be implemented in routers therefore we investigated 4 new services to improve the performances of local recovery and heterogeneity support. These are:
the dynamic replier election which consists of choosing a link/host as a replier one (the one which will send again the missing packet) to perform local recoveries;
the early detection of packet losses and the emission of the corresponding NACKs;
an accurate, on a per-hop basis, RTT (Round-Trip-Time) computation for congestion and rate adaptation purposes for interoperability with unicast TCP flows;
the partitioning of the receivers to handle heterogeneity.
For the first two services, we have conducted analytical studies similar to those realized in 2001 [73] to model and evaluate their performances. The results have been published in [76] and [74] , [14] [15] . Regarding the two last services consisting in the RTTs aggregation and the receiver partitioning, preliminary results published in [37] , [36] and [38] are very encouraging.
All the proposed services are lightweight active services that consume very few router's resources. However, combined with local recoveries, they are very beneficial to reduce the end-to-end latency and to provide the support of heterogeneity in a multicast session.
Key words: Reliable multicast, programmable networks, congestion control.
Congestion control in multicast is a difficult task because it is hard to get and take into account the status of the entire group of receivers and to satisfy all the receivers when they are heterogeneous (which is almost always the case).
The active service that estimates the RTTs from the receivers towards the source is an important component in the congestion control mechanism that we proposed. Active routers in the multicast tree estimate the RTT towards their parent node (another active router or the source) and aggregate these informations in order to propagate only one value towards the source. As in RMANP [54] or NCA [68] , we benefit from the physical multicast tree to aggregate the RTT values, as opposed to TRAM [58] or MTCP [81] which use, and thus maintain, a logical tree.
The AMCA algorithm (Active-based Multicast Congestion Avoidance Algorithm) [37] that we proposed use this lightweight service to predict (and in most cases avoid) congestions by observing the RTT variation. The approach is similar in concepts to TCP Vegas but do not suffer from the path re-routing problem. AMCA is compatible with TCP.
Key words: Reliable multicast, programmable networks, protocol.
We have integrated in a protocol called DyRAM (Dynamic Replier Active reliable Multicast), the active services that we proposed (along with feedback aggregation and the subcast feature). The main objective of DyRAM is to avoid cache in routers and to provide low recovery latencies. DyRAM is therefore very different from ARM [71] , AER [68] or MAF [82] . DyRAM and its performance are described in [77] and [19] .
|
|
|