[ensembl-dev] Performance Analysis and Hardware Acceleration of VEP.

[Syed Hossain]

Hello Muhammad Ali Akhtar,

The article had the latest performance analysis for VEP. It is indeed 
not an up-to-date analysis but may give you some idea. We do have plan 
to do a performance analysis soon and give up-to-date result.

You can look at code implementation here - 
https://github.com/Ensembl/ensembl-variation/tree/release/107/C_code 
(VEP uses ensembl-variation APIs for various bits).

Yes, I would say VEP is not optimised to run on FPGA/GPUs. It is 
interesting area to explore but unfortunately I cannot comment on a 
timeline where we will be working on this and have some tangible 
outcome.

Best regards,
Nakib

On 2022-08-30 19:30, Muhammad Ali Akhtar wrote:
> Dear Syed Hossain,
> 
> Thanks for the response. The article you referred to was published in
> 2016. There have been many vep releases after that and I think a lot
> has changed. Are you saying that the runtime / performance numbers
> cited in that article in 2016 are still valid?
> 
> Secondly, you mentioned one underlying computation step (  linkage
> disequilibrium stats ). Can you point me to any publication / study
> that lists the algorithms (and mathematical operations ) used in VEP
> or other variant analysis tools? If there are no studies, can you
> atleast point me to perl files that implement   linkage disequilibrium
> stats  step.
> 
> Third: I read about the fork  / multithreaded option. However, this
> option is still meant for CPU execution i.e. Multi-Core CPU execution.
> We can create many (100-200) simple cores (PEs) in a mid-sized FPGA
> chip. However, these simple cores will not be as complex as Xeon
> Processor cores. e.g. FPGA cores are not good for Floating Point Ops.
> 
> I was just wondering if there are any mathematical operations in VEP
> algorithms that can only be run on CPUs. In other words, there are
> steps in VEP processing that are not suitable for GPUs / FPGA chips.
> FPGA chips are better at simple Integer Arithmetic and logical
> operations and if mathematical operations in VEP are also simple
> integer arithmetic and logical operations, 100 cores in FPGA can
> theoretically give 100x Speed Up over One Xeon Core.
> 
> Muhammad Ali Akhtar
> 
> On Tue, Aug 30, 2022 at 1:52 PM Syed Hossain <snhossain at ebi.ac.uk>
> wrote:
> 
>> Hello Muhammad Ali Akhtar,
>> 
>> Thanks for the query!
>> 
>> 1. So we do have our performance analysis on VEP which was published
>> on
>> the latest paper. You can check the result here -
>> 
> https://genomebiology.biomedcentral.com/articles/10.1186/s13059-016-0974-4.
>> 
>> 2. VEP have option to run in multiple thread (--fork option, see -
>> 
> https://www.ensembl.org/info/docs/tools/vep/script/vep_options.html#opt_fork)
>> 
>> which suitable to run on multiple core clusters. And some of the
>> underlying computation (such as linkage disequilibrium stats) are in
>> C
>> and Perl works as a wrapper. We do have a longer time plan to switch
>> to
>> a modern language.
>> 
>> Hope that answers your query.
>> 
>> Best regards,
>> Nakib
>> 
>> On 2022-08-27 14:20, Muhammad Ali Akhtar wrote:
>>> Hello Everyone,
>>> 
>>> This is my first time interacting with the VEP dev team. Just had
>> a
>>> few questions.
>>> 
>>> 1. Are there any studies / reports out there with runtime /
>>> performance analysis of VEP?
>>> 2. Have there been any efforts on improving VEP performance? like
>>> coding in different languages (C/C++/) or using different
>> processors
>>> (GPUs / FPGAs)?
>>> 
>>> I ask this because many steps (Like sequence alignment, variant
>>> calling)  in the genome processing pipeline are being targeted for
>>> hardware acceleration but I couldn't find anything on customising
>> /
>>> improving VEP run time. Is it because the runtime of VEP has been
>>> short / average even for most demanding workloads?
>>> 
>>> Regards,
>>> Muhammad Ali Akhtar
>>> _______________________________________________
>>> Dev mailing list    Dev at ensembl.org
>>> Posting guidelines and subscribe/unsubscribe info:
>>> https://lists.ensembl.org/mailman/listinfo/dev_ensembl.org
>>> Ensembl Blog: http://www.ensembl.info/