- Ahmed Farid added an answer:Best method for a wired network bus for scalable SoC clusters (Using Raspberry Pi - Compute Module)?
Taking the Raspberry Pi Compute Module as an example, an interesting use case is to build a scalable computing system that utilizes each module as a cluster.
However I am a little curious on the best wired networking model here. Should it be a master-slave design using I2C for example (I don't think its reliable or quick enough)? What are your thoughts here?
Thank you all for your responses.
I have seen a lot of RPi cluster projects and all used ethernet as a networking interface indeed, but is this the most suited for SoCs like the recently released RPi Compute Modules that only communicate via GPIO?
NoCs start to make a lot of sense here indeed, rather than connecting a wire per cluster and the need to obtain a new router whenever all slots are occupied. NoC approach would help a lot in conserving such resources and reducing their scaling as we scale up clusters. However this approach from my humble knowledge would focus mainly on designing a special board and more localized networking, which brings me to the proposed use of a master-slave approach.
A master would be the main point of contact for receiving commands and scheduling them across the other clusters (Slaves), then finally consolidating the result sets. I believe something like this was definitely made before.
@Barry I'd love to join further discussions as well! I'm recently shifting my technical interests towards distributed systems and preparing for research as well.Following
- Abdallah Sayyed-Ahmad added an answer:What is the meaning of "fullSamples" keyword in ABF calculation in NAMD?
While doing Adaptive Biasing Force calculations in NAMD we need to provide the keyword "fullSamples" the default value of which is 200. What is the significance of this keyword? How do I provide the keyword according to the needs of my system?
It is the number of phase space points needed to adequately estimate the average biasing force. Obviously the more points you can use the better, but due to limited time and computational resources one has to find "what is good enough number of points" This can be done by starting with the default number and then incrementally increase it until a reasonable estimate of the average biasing force is obtained.Following
- Qasem Abu Al-Haija added an answer:How to build or develop an optimal mathematical model for the scalability of the WSN.Many Parameters affects the Network Scalability; Throughput, QoS, Network Size, Routing, etc.Here is the paperFollowing
- Augusto Ciuffoletti added an answer:How do we define and quantify scalability, elasticity and efficiency?These new quality attributes arise especially for Cloud Computing applications.I think that elasticity is the property that distinguishes cloud computing from other computing paradigms. A metric for elasticity might be relevant for cloud users: how fast the system responds to your demand? This is an issue for major providers, since the response time must take into account the time needed to find the resource, allocate, configure and deliver it to the user. In conclusion, the time needed to allocate 100 CPUs may be higher or lower than one hundred times the time needed to allocate just one: this depends on many cloud management decisions. The user might appreciate a metric of this sort.
Anther issue related to this aspect is how the cloud reacts to a decrease in the demand: while it is clear that if you explicitely deallocate a node, the effective deallocation should occur sooner or later, it is not clear how cloud management responds if the the user subscribed an "adaptive" service. In both cases, a slow response has a financial impact.
To come to your question, elasticity should be related to the timeliness with which the cloud reacts to a changing demand: both positive and negative changes. A metric should take into account the time needed to effect the change: cpu/sec is the rough idea for computing resources. Financial and management issues (need to change the contract?) contribute to the figure.Following
- Lucian Hancu added an answer:Can someone suggest some research aspects of scalability in cloud?Want to know internal mechanism of cloud auto scaling@ Pavan: the easiest way of learning is by experimenting.Following
- Carlos Lübbe added an answer:What are the most common practices in providing high scale access to spatial data with region query support?An increasing demand for geographic data imposes high query loads on data providers. Main players in the web (e.g. Facebook) use distributed memory caching systems such as Memcached to cache frequently accessed data in a key-value fashion. However, for spatial/geographic data as provided by OpenStreetMap for instance, the key-value paradigm is insufficient or at least unhandy, as many location-based services or applications require data which is located in certain ranges or regions.
How do service providers most commonly implement high scale access to spatial data with region query support?
(a) Is there any widely used distributed caching system similar to Memcached available, which supports region queries out of the box?
(b) Do application developers use standard key-value systems (such as Memcached) and manually remap query ranges to keys via space-filling curves or other spatial re-mapping techniques.
(c) Or can scalability be provided by a relational database with spatial extender which is extensively partitioned and replicated among several db instances.Hi Javier and Ron,
thanks for the reference to Django, an ORM I have not been aware of until now. Interesting enough is that it also provides a key-value cache API using Memcached as back-end. And that takes me back to Ron's space-filling curves which are required to cache spatial data in key-value fashion. So thanks, Ron, for your offer. I will probably come back to it, as soon as I have started the coding part of this matter.Following
- Eduardo Sanchez added an answer:Has anybody tried to solve a banded system of equation using ScaLAPACK from a C code, attaining scalable results?Further details: http://www.linuxquestions.org/questions/showthread.php?p=4897114#post4897114Dear Theodore, thanks for your reply, which makes perfect sense. MTL4 is written in C++, so I would try that last. A highly optimized Intel MKL, can't be, since my advisors don't want to spend the money on it. NOW, my question basically is: regarless of the fact that ScaLAPACK, as downloaded from NetLib, is old, it certainly gives me scalable results. So why can't I find a C code, calling it, to solve a banded system? I also want to try SPIKE, which has not been my first choice, simply because my advisors are allergic to proprietary code.Following