Paper Summary: Occupy The Cloud: Distributed Computing For The 99%
"We are the 99%!" (Occupy Wall Street Movement, 2011)
The 99% that the championship of this newspaper refers to is the non-cloud-native in addition to non-CS-native programmers. Most scientific in addition to analytics software is written yesteryear domain experts similar biologists in addition to physicists rather than calculator scientists. Writing in addition to deploying at the cloud is difficult for these folks. Heck it is fifty-fifty difficult for the calculator scientific discipline folk. The newspaper reports that an informal survey at UC Berkeley flora that the bulk of auto learning graduate students stimulate got never written a cluster computing undertaking due to complexity of setting upward cloud platforms.
Yes, cloud computing virtualized a lot of things, in addition to VMs, in addition to late containers, reduced the friction of deploying at the clouds. However, at that spot are soundless also many choices to brand in addition to things to configure before y'all tin dismiss acquire your code to deploy & run at the cloud. We soundless don't stimulate got a "cloud button", where y'all tin dismiss force to acquire your unmarried auto code deployed in addition to running on the cloud inwards seconds.
But nosotros are getting there. AWS Lambda in addition to Google Cloud Functions aim to solve this occupation yesteryear providing infrastructure to run event-driven, stateless functions every bit microservices. In this "serverless" model, a role is deployed ane time in addition to is invoked repeatedly whenever novel inputs arrive. Thus the serverless model elastically scales amongst input size. Here is an before postal service from me summarizing a newspaper on the serverless computation model.
This paper, which appeared on arXiv inwards Feb 2017 in addition to revised June 2017, pushes the envelope on the serverless model farther inwards social club to implement distributed information processing in addition to analytics applications. The newspaper is a vision paper, thus it is depression on details at roughly parts, nevertheless a image system, PyWren, developed inwards Python over AWS Lambda, is made available every bit opensource.
In social club to construct a information processing system, the newspaper dynamically injects code into these stateless AWS Lambda functions to circumvent its limits in addition to extend its capabilities. The model has ane unproblematic primitive: users submit functions that are executed inwards a remote container; the functions are stateless; the soil every bit good every bit input, in addition to output is relegated to the shared remote storage. (This fits good amongst the ascension tendency of the disaggregated storage architecture.) Surprisingly, the newspaper finds that the functioning degradation from using such an approach is negligible for many workloads.
After summarizing the code injection approach, I volition shout out how PyWren tin dismiss implement increasingly to a greater extent than sophisticated information processing applications ranging from all Map, to Map in addition to monolithic reduce, in addition to MapReduce, in addition to finally a hint of parameter-server implementation.
PyWren serializes the user submitted Python role using cloudpickle. PyWren submits the serialized role along amongst each serialized datum yesteryear placing them into globally unique keys inwards S3, in addition to thus invokes a mutual Lambda function. On the server side, PyWren invokes the relevant role on the relevant datum, both extracted from S3. The termination of the role invocation is serialized in addition to placed dorsum into S3 at a pre-specified key, in addition to undertaking completion is signaled yesteryear the beingness of this key. In this way, PyWren is able to reuse ane registered Lambda role to execute unlike user Python functions in addition to mitigate the high latency for role registration, spell executing functions that order off Lambda’s code size limit.
Map + Monolithic reduce. An piece of cake agency to implement MapReduce is to practise the Reduce inwards simply ane machine. For this ane auto to perform reduce, they usage a dedicated unmarried r4.16xlarge instance. This auto offers a rattling large amount of CPU in addition to RAM for $14 an hour.
MapReduce via BSP. To perform Reduce over many workers, nosotros tin dismiss usage the mass synchronous processing (BSP) model. To implement the BSP model in addition to information shuffling across the stages PyWren leverages the high-bandwidth remote storage AWS S3 provides. To showcase this approach, they implemented a give-and-take count computer programme inwards PyWren in addition to flora that on 83M items, it is alone 17% slower than PySpark running on dedicated servers.
The newspaper does non pull how BSP is implemented. I gauge this is the responsibleness of the driver computer programme on the scientist's laptop. Eric Jonas, ane of the authors of this work, calls this the shim handler, that submits the lambda functions to AWS. So I gauge this driver checks the progress on the rounds yesteryear polling S3, in addition to prepare/invoke the lambda functions for the side yesteryear side round.
The newspaper also implements a to a greater extent than ambitious application, Terasort, using the PyWren MapReduce. Since this application produces a lot of intermediate files to shuffle inwards between, they say S3 becomes a bottleneck. So they usage AWS elastic cache, a Redis in-memory key-value store. Using this, they demo that PyWren tin dismiss course of study 1TB information inwards 3.4 minutes using thou lambda workers.
The Parameter-server implementation. The newspaper claims to also implement Parameter-Server ane time to a greater extent than using Redis inmemory keyvalue store. But at that spot are no details, thus it is unclear if the functioning of using that is acceptable.
Using AWS Lambda is alone 2× to a greater extent than expensive than on-demand instances. The newspaper says that this terms is worthwhile "given substantially finer-grained billing, much greater elasticity, in addition to the fact that many dedicated clusters are often running at 50% utilization".
As for limitations, this industrial plant best if the workers practise non demand to coordinate oftentimes in addition to usage around of the v minutes (i.e. 300s) of lambda role execution fourth dimension for computing over the information input to its 1.5GB RAM to hit the output data. So the newspaper cautions that for applications similar particle simulations, which require a lot of coordination betwixt long running processes, the PyWren model of using stateless functions amongst remote storage powerfulness non live a goodness fit.
It looks similar beyond the map functionality, relaxation of programming is soundless non that great. But this is a measuring inwards the correct direction.
The 99% that the championship of this newspaper refers to is the non-cloud-native in addition to non-CS-native programmers. Most scientific in addition to analytics software is written yesteryear domain experts similar biologists in addition to physicists rather than calculator scientists. Writing in addition to deploying at the cloud is difficult for these folks. Heck it is fifty-fifty difficult for the calculator scientific discipline folk. The newspaper reports that an informal survey at UC Berkeley flora that the bulk of auto learning graduate students stimulate got never written a cluster computing undertaking due to complexity of setting upward cloud platforms.
Yes, cloud computing virtualized a lot of things, in addition to VMs, in addition to late containers, reduced the friction of deploying at the clouds. However, at that spot are soundless also many choices to brand in addition to things to configure before y'all tin dismiss acquire your code to deploy & run at the cloud. We soundless don't stimulate got a "cloud button", where y'all tin dismiss force to acquire your unmarried auto code deployed in addition to running on the cloud inwards seconds.
But nosotros are getting there. AWS Lambda in addition to Google Cloud Functions aim to solve this occupation yesteryear providing infrastructure to run event-driven, stateless functions every bit microservices. In this "serverless" model, a role is deployed ane time in addition to is invoked repeatedly whenever novel inputs arrive. Thus the serverless model elastically scales amongst input size. Here is an before postal service from me summarizing a newspaper on the serverless computation model.
This paper, which appeared on arXiv inwards Feb 2017 in addition to revised June 2017, pushes the envelope on the serverless model farther inwards social club to implement distributed information processing in addition to analytics applications. The newspaper is a vision paper, thus it is depression on details at roughly parts, nevertheless a image system, PyWren, developed inwards Python over AWS Lambda, is made available every bit opensource.
In social club to construct a information processing system, the newspaper dynamically injects code into these stateless AWS Lambda functions to circumvent its limits in addition to extend its capabilities. The model has ane unproblematic primitive: users submit functions that are executed inwards a remote container; the functions are stateless; the soil every bit good every bit input, in addition to output is relegated to the shared remote storage. (This fits good amongst the ascension tendency of the disaggregated storage architecture.) Surprisingly, the newspaper finds that the functioning degradation from using such an approach is negligible for many workloads.
After summarizing the code injection approach, I volition shout out how PyWren tin dismiss implement increasingly to a greater extent than sophisticated information processing applications ranging from all Map, to Map in addition to monolithic reduce, in addition to MapReduce, in addition to finally a hint of parameter-server implementation.
Code injection
An AWS Lambda role gives y'all v minutes (300secs) of execution fourth dimension at unmarried total in addition to 1.5 Gb RAM, in addition to also gives y'all 512 MB inwards /tmp. PyWren exploits this 512MB tmp infinite to read Anaconda Python Runtime libraries. (This linked verbalise clarified PyWren code injection for me.)
PyWren serializes the user submitted Python role using cloudpickle. PyWren submits the serialized role along amongst each serialized datum yesteryear placing them into globally unique keys inwards S3, in addition to thus invokes a mutual Lambda function. On the server side, PyWren invokes the relevant role on the relevant datum, both extracted from S3. The termination of the role invocation is serialized in addition to placed dorsum into S3 at a pre-specified key, in addition to undertaking completion is signaled yesteryear the beingness of this key. In this way, PyWren is able to reuse ane registered Lambda role to execute unlike user Python functions in addition to mitigate the high latency for role registration, spell executing functions that order off Lambda’s code size limit.
Implementing Map, MapReduce, in addition to the Parameter-Server
Map implementation. Many scientific in addition to analytic workloads are embarrassingly parallel. The map primitive provided yesteryear PyWren makes addressing these usage cases easy. Calling the map launches every bit many stateless functions every bit at that spot are elements inwards the listing that ane is mapping over.Map + Monolithic reduce. An piece of cake agency to implement MapReduce is to practise the Reduce inwards simply ane machine. For this ane auto to perform reduce, they usage a dedicated unmarried r4.16xlarge instance. This auto offers a rattling large amount of CPU in addition to RAM for $14 an hour.
MapReduce via BSP. To perform Reduce over many workers, nosotros tin dismiss usage the mass synchronous processing (BSP) model. To implement the BSP model in addition to information shuffling across the stages PyWren leverages the high-bandwidth remote storage AWS S3 provides. To showcase this approach, they implemented a give-and-take count computer programme inwards PyWren in addition to flora that on 83M items, it is alone 17% slower than PySpark running on dedicated servers.
The newspaper does non pull how BSP is implemented. I gauge this is the responsibleness of the driver computer programme on the scientist's laptop. Eric Jonas, ane of the authors of this work, calls this the shim handler, that submits the lambda functions to AWS. So I gauge this driver checks the progress on the rounds yesteryear polling S3, in addition to prepare/invoke the lambda functions for the side yesteryear side round.
The newspaper also implements a to a greater extent than ambitious application, Terasort, using the PyWren MapReduce. Since this application produces a lot of intermediate files to shuffle inwards between, they say S3 becomes a bottleneck. So they usage AWS elastic cache, a Redis in-memory key-value store. Using this, they demo that PyWren tin dismiss course of study 1TB information inwards 3.4 minutes using thou lambda workers.
The Parameter-server implementation. The newspaper claims to also implement Parameter-Server ane time to a greater extent than using Redis inmemory keyvalue store. But at that spot are no details, thus it is unclear if the functioning of using that is acceptable.
Discussion
They detect that it is possible to accomplish around 30-40 MB/s write in addition to read functioning per total to S3, matching the per-core functioning of a unmarried local SSD on typical EC2 nodes. They also demo that this scales to 60-80 GB/s to S3 across 2800 simultaneous functions.Using AWS Lambda is alone 2× to a greater extent than expensive than on-demand instances. The newspaper says that this terms is worthwhile "given substantially finer-grained billing, much greater elasticity, in addition to the fact that many dedicated clusters are often running at 50% utilization".
As for limitations, this industrial plant best if the workers practise non demand to coordinate oftentimes in addition to usage around of the v minutes (i.e. 300s) of lambda role execution fourth dimension for computing over the information input to its 1.5GB RAM to hit the output data. So the newspaper cautions that for applications similar particle simulations, which require a lot of coordination betwixt long running processes, the PyWren model of using stateless functions amongst remote storage powerfulness non live a goodness fit.
It looks similar beyond the map functionality, relaxation of programming is soundless non that great. But this is a measuring inwards the correct direction.
0 Response to "Paper Summary: Occupy The Cloud: Distributed Computing For The 99%"
Post a Comment