Re: [Rd] [parallel] fixes load balancing of parLapplyLB

[Tomas Kalibera]

Chunk size support has been added in R-devel 74353. Please let me know 
if you find any problem.


Thanks,
Tomas

On 03/01/2018 09:19 AM, Christian Krause wrote:

Dear Tomas,

Thanks for your commitment to fix this issue and also to add the chunk size as 
an argument. If you want our input, let us know ;)

Best Regards

On 02/26/2018 04:01 PM, Tomas Kalibera wrote:

Dear Christian and Henrik,

thank you for spotting the problem and suggestions for a fix. We'll probably 
add a chunk.size argument to parLapplyLB and parLapply to follow OpenMP 
terminology, which has already been an inspiration for the present code 
(parLapply already implements static scheduling via internal function 
staticClusterApply, yet with a fixed chunk size; parLapplyLB already implements 
dynamic scheduling via internal function dynamicClusterApply, but with a fixed 
chunk size set to an unlucky value so that it behaves like static scheduling). 
The default chunk size for parallelLapplyLB will be set so that there is some 
dynamism in the schedule even by default. I am now testing a patch with these 
changes.

Best
Tomas


On 02/20/2018 11:45 AM, Christian Krause wrote:

Dear Henrik,

The rationale is just that it is within these extremes and that it is really 
simple to calculate, without making any assumptions and knowing that it won't 
be perfect.

The extremes A and B you are mentioning are special cases based on assumptions. 
Case A is based on the assumption that the function has a long runtime or 
varying runtime, then you are likely to get the best load balancing with really 
small chunks. Case B is based on the assumption that the function runtime is 
the same for each list element, i.e. where you don't actually need load 
balancing, i.e. just use `parLapply` without load balancing.

This new default is **not the best one**. It's just a better one than we had 
before. There is no best one we can use as default because **we don't know the 
function runtime and how it varies**. The user needs to decide that because 
he/she knows the function. As mentioned before, I will write a patch that makes 
the chunk size an optional argument, so the user can decide because only he/she 
has all the information to choose the best chunk size, just like you did with 
the `future.scheduling` parameter.

Best Regards

On February 19, 2018 10:11:04 PM GMT+01:00, Henrik Bengtsson 
 wrote:

Hi, I'm trying to understand the rationale for your proposed amount of
splitting and more precisely why that one is THE one.

If I put labels on your example numbers in one of your previous post:

nbrOfElements <- 97
nbrOfWorkers <- 5

With these, there are two extremes in how you can split up the
processing in chunks such that all workers are utilized:

(A) Each worker, called multiple times, processes one element each
time:


nbrOfElements <- 97
nbrOfWorkers <- 5
nbrOfChunks <- nbrOfElements
sapply(parallel:::splitList(1:nbrOfElements, nbrOfChunks), length)

[1] 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
[30] 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
[59] 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
[88] 1 1 1 1 1 1 1 1 1 1


(B) Each worker, called once, processes multiple element:


nbrOfElements <- 97
nbrOfWorkers <- 5
nbrOfChunks <- nbrOfWorkers
sapply(parallel:::splitList(1:nbrOfElements, nbrOfChunks), length)

[1] 20 19 19 19 20

I understand that neither of these two extremes may be the best when
it comes to orchestration overhead and load balancing. Instead, the
best might be somewhere in-between, e.g.

(C) Each worker, called multiple times, processing multiple elements:


nbrOfElements <- 97
nbrOfWorkers <- 5
nbrOfChunks <- nbrOfElements / nbrOfWorkers
sapply(parallel:::splitList(1:nbrOfElements, nbrOfChunks), length)

[1] 5 5 5 5 4 5 5 5 5 5 4 5 5 5 5 4 5 5 5 5

However, there are multiple alternatives between the two extremes, e.g.


nbrOfChunks <- scale * nbrOfElements / nbrOfWorkers

So, is there a reason why you argue for scale = 1.0 to be the optimal?

FYI, In future.apply::future_lapply(X, FUN, ...) there is a
'future.scheduling' scale factor(*) argument where default
future.scheduling = 1 corresponds to (B) and future.scheduling = +Inf
to (A).  Using future.scheduling = 4 achieves the amount of
load-balancing you propose in (C).   (*) Different definition from the
above 'scale'. (Disclaimer: I'm the author)

/Henrik

On Mon, Feb 19, 2018 at 10:21 AM, Christian Krause
 wrote:

Dear R-Devel List,

I have installed R 3.4.3 with the patch applied on our cluster and

ran a *real-world* job of one of our users to confirm that the patch
works to my satisfaction. Here are the results.

The original was a series of jobs, all essentially doing the same

stuff using bootstrapped data, so for the original there is more data
and I show the arithmetic mean with standard deviation. The
confirmation with the patched R was only a single instance of that

Re: [Rd] [parallel] fixes load balancing of parLapplyLB

[Christian Krause]

Dear Tomas,

Thanks for your commitment to fix this issue and also to add the chunk size as 
an argument. If you want our input, let us know ;)

Best Regards

On 02/26/2018 04:01 PM, Tomas Kalibera wrote:
> Dear Christian and Henrik,
> 
> thank you for spotting the problem and suggestions for a fix. We'll probably 
> add a chunk.size argument to parLapplyLB and parLapply to follow OpenMP 
> terminology, which has already been an inspiration for the present code 
> (parLapply already implements static scheduling via internal function 
> staticClusterApply, yet with a fixed chunk size; parLapplyLB already 
> implements dynamic scheduling via internal function dynamicClusterApply, but 
> with a fixed chunk size set to an unlucky value so that it behaves like 
> static scheduling). The default chunk size for parallelLapplyLB will be set 
> so that there is some dynamism in the schedule even by default. I am now 
> testing a patch with these changes.
> 
> Best
> Tomas
> 
> 
> On 02/20/2018 11:45 AM, Christian Krause wrote:
>> Dear Henrik,
>>
>> The rationale is just that it is within these extremes and that it is really 
>> simple to calculate, without making any assumptions and knowing that it 
>> won't be perfect.
>>
>> The extremes A and B you are mentioning are special cases based on 
>> assumptions. Case A is based on the assumption that the function has a long 
>> runtime or varying runtime, then you are likely to get the best load 
>> balancing with really small chunks. Case B is based on the assumption that 
>> the function runtime is the same for each list element, i.e. where you don't 
>> actually need load balancing, i.e. just use `parLapply` without load 
>> balancing.
>>
>> This new default is **not the best one**. It's just a better one than we had 
>> before. There is no best one we can use as default because **we don't know 
>> the function runtime and how it varies**. The user needs to decide that 
>> because he/she knows the function. As mentioned before, I will write a patch 
>> that makes the chunk size an optional argument, so the user can decide 
>> because only he/she has all the information to choose the best chunk size, 
>> just like you did with the `future.scheduling` parameter.
>>
>> Best Regards
>>
>> On February 19, 2018 10:11:04 PM GMT+01:00, Henrik Bengtsson 
>>  wrote:
>>> Hi, I'm trying to understand the rationale for your proposed amount of
>>> splitting and more precisely why that one is THE one.
>>>
>>> If I put labels on your example numbers in one of your previous post:
>>>
>>> nbrOfElements <- 97
>>> nbrOfWorkers <- 5
>>>
>>> With these, there are two extremes in how you can split up the
>>> processing in chunks such that all workers are utilized:
>>>
>>> (A) Each worker, called multiple times, processes one element each
>>> time:
>>>
 nbrOfElements <- 97
 nbrOfWorkers <- 5
 nbrOfChunks <- nbrOfElements
 sapply(parallel:::splitList(1:nbrOfElements, nbrOfChunks), length)
>>> [1] 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
>>> [30] 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
>>> [59] 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
>>> [88] 1 1 1 1 1 1 1 1 1 1
>>>
>>>
>>> (B) Each worker, called once, processes multiple element:
>>>
 nbrOfElements <- 97
 nbrOfWorkers <- 5
 nbrOfChunks <- nbrOfWorkers
 sapply(parallel:::splitList(1:nbrOfElements, nbrOfChunks), length)
>>> [1] 20 19 19 19 20
>>>
>>> I understand that neither of these two extremes may be the best when
>>> it comes to orchestration overhead and load balancing. Instead, the
>>> best might be somewhere in-between, e.g.
>>>
>>> (C) Each worker, called multiple times, processing multiple elements:
>>>
 nbrOfElements <- 97
 nbrOfWorkers <- 5
 nbrOfChunks <- nbrOfElements / nbrOfWorkers
 sapply(parallel:::splitList(1:nbrOfElements, nbrOfChunks), length)
>>> [1] 5 5 5 5 4 5 5 5 5 5 4 5 5 5 5 4 5 5 5 5
>>>
>>> However, there are multiple alternatives between the two extremes, e.g.
>>>
 nbrOfChunks <- scale * nbrOfElements / nbrOfWorkers
>>> So, is there a reason why you argue for scale = 1.0 to be the optimal?
>>>
>>> FYI, In future.apply::future_lapply(X, FUN, ...) there is a
>>> 'future.scheduling' scale factor(*) argument where default
>>> future.scheduling = 1 corresponds to (B) and future.scheduling = +Inf
>>> to (A).  Using future.scheduling = 4 achieves the amount of
>>> load-balancing you propose in (C).   (*) Different definition from the
>>> above 'scale'. (Disclaimer: I'm the author)
>>>
>>> /Henrik
>>>
>>> On Mon, Feb 19, 2018 at 10:21 AM, Christian Krause
>>>  wrote:
 Dear R-Devel List,

 I have installed R 3.4.3 with the patch applied on our cluster and
>>> ran a *real-world* job of one of our users to confirm that the patch
>>> works to my satisfaction. Here are the results.
 The original was a series of jobs, all essentially doing the same
>>> 

Re: [Rd] [parallel] fixes load balancing of parLapplyLB

[Tomas Kalibera]

Dear Christian and Henrik,

thank you for spotting the problem and suggestions for a fix. We'll 
probably add a chunk.size argument to parLapplyLB and parLapply to 
follow OpenMP terminology, which has already been an inspiration for the 
present code (parLapply already implements static scheduling via 
internal function staticClusterApply, yet with a fixed chunk size; 
parLapplyLB already implements dynamic scheduling via internal function 
dynamicClusterApply, but with a fixed chunk size set to an unlucky value 
so that it behaves like static scheduling). The default chunk size for 
parallelLapplyLB will be set so that there is some dynamism in the 
schedule even by default. I am now testing a patch with these changes.


Best
Tomas


On 02/20/2018 11:45 AM, Christian Krause wrote:

Dear Henrik,

The rationale is just that it is within these extremes and that it is really 
simple to calculate, without making any assumptions and knowing that it won't 
be perfect.

The extremes A and B you are mentioning are special cases based on assumptions. 
Case A is based on the assumption that the function has a long runtime or 
varying runtime, then you are likely to get the best load balancing with really 
small chunks. Case B is based on the assumption that the function runtime is 
the same for each list element, i.e. where you don't actually need load 
balancing, i.e. just use `parLapply` without load balancing.

This new default is **not the best one**. It's just a better one than we had 
before. There is no best one we can use as default because **we don't know the 
function runtime and how it varies**. The user needs to decide that because 
he/she knows the function. As mentioned before, I will write a patch that makes 
the chunk size an optional argument, so the user can decide because only he/she 
has all the information to choose the best chunk size, just like you did with 
the `future.scheduling` parameter.

Best Regards

On February 19, 2018 10:11:04 PM GMT+01:00, Henrik Bengtsson 
 wrote:

Hi, I'm trying to understand the rationale for your proposed amount of
splitting and more precisely why that one is THE one.

If I put labels on your example numbers in one of your previous post:

nbrOfElements <- 97
nbrOfWorkers <- 5

With these, there are two extremes in how you can split up the
processing in chunks such that all workers are utilized:

(A) Each worker, called multiple times, processes one element each
time:


nbrOfElements <- 97
nbrOfWorkers <- 5
nbrOfChunks <- nbrOfElements
sapply(parallel:::splitList(1:nbrOfElements, nbrOfChunks), length)

[1] 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
[30] 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
[59] 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
[88] 1 1 1 1 1 1 1 1 1 1


(B) Each worker, called once, processes multiple element:


nbrOfElements <- 97
nbrOfWorkers <- 5
nbrOfChunks <- nbrOfWorkers
sapply(parallel:::splitList(1:nbrOfElements, nbrOfChunks), length)

[1] 20 19 19 19 20

I understand that neither of these two extremes may be the best when
it comes to orchestration overhead and load balancing. Instead, the
best might be somewhere in-between, e.g.

(C) Each worker, called multiple times, processing multiple elements:


nbrOfElements <- 97
nbrOfWorkers <- 5
nbrOfChunks <- nbrOfElements / nbrOfWorkers
sapply(parallel:::splitList(1:nbrOfElements, nbrOfChunks), length)

[1] 5 5 5 5 4 5 5 5 5 5 4 5 5 5 5 4 5 5 5 5

However, there are multiple alternatives between the two extremes, e.g.


nbrOfChunks <- scale * nbrOfElements / nbrOfWorkers

So, is there a reason why you argue for scale = 1.0 to be the optimal?

FYI, In future.apply::future_lapply(X, FUN, ...) there is a
'future.scheduling' scale factor(*) argument where default
future.scheduling = 1 corresponds to (B) and future.scheduling = +Inf
to (A).  Using future.scheduling = 4 achieves the amount of
load-balancing you propose in (C).   (*) Different definition from the
above 'scale'. (Disclaimer: I'm the author)

/Henrik

On Mon, Feb 19, 2018 at 10:21 AM, Christian Krause
 wrote:

Dear R-Devel List,

I have installed R 3.4.3 with the patch applied on our cluster and

ran a *real-world* job of one of our users to confirm that the patch
works to my satisfaction. Here are the results.

The original was a series of jobs, all essentially doing the same

stuff using bootstrapped data, so for the original there is more data
and I show the arithmetic mean with standard deviation. The
confirmation with the patched R was only a single instance of that
series of jobs.

## Job Efficiency

The job efficiency is defined as (this is what the `qacct-efficiency`

tool below does):

```
efficiency = cputime / cores / wallclocktime * 100%
```

In simpler words: how well did the job utilize its CPU cores. It

shows the percentage of time the job was actually doing stuff, as
opposed to the difference:

```
wasted = 100% - 

Re: [Rd] [parallel] fixes load balancing of parLapplyLB

[Christian Krause]

Dear Henrik,

The rationale is just that it is within these extremes and that it is really 
simple to calculate, without making any assumptions and knowing that it won't 
be perfect.

The extremes A and B you are mentioning are special cases based on assumptions. 
Case A is based on the assumption that the function has a long runtime or 
varying runtime, then you are likely to get the best load balancing with really 
small chunks. Case B is based on the assumption that the function runtime is 
the same for each list element, i.e. where you don't actually need load 
balancing, i.e. just use `parLapply` without load balancing.

This new default is **not the best one**. It's just a better one than we had 
before. There is no best one we can use as default because **we don't know the 
function runtime and how it varies**. The user needs to decide that because 
he/she knows the function. As mentioned before, I will write a patch that makes 
the chunk size an optional argument, so the user can decide because only he/she 
has all the information to choose the best chunk size, just like you did with 
the `future.scheduling` parameter.

Best Regards

On February 19, 2018 10:11:04 PM GMT+01:00, Henrik Bengtsson 
 wrote:
>Hi, I'm trying to understand the rationale for your proposed amount of
>splitting and more precisely why that one is THE one.
>
>If I put labels on your example numbers in one of your previous post:
>
> nbrOfElements <- 97
> nbrOfWorkers <- 5
>
>With these, there are two extremes in how you can split up the
>processing in chunks such that all workers are utilized:
>
>(A) Each worker, called multiple times, processes one element each
>time:
>
>> nbrOfElements <- 97
>> nbrOfWorkers <- 5
>> nbrOfChunks <- nbrOfElements
>> sapply(parallel:::splitList(1:nbrOfElements, nbrOfChunks), length)
> [1] 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
>[30] 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
>[59] 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
>[88] 1 1 1 1 1 1 1 1 1 1
>
>
>(B) Each worker, called once, processes multiple element:
>
>> nbrOfElements <- 97
>> nbrOfWorkers <- 5
>> nbrOfChunks <- nbrOfWorkers
>> sapply(parallel:::splitList(1:nbrOfElements, nbrOfChunks), length)
>[1] 20 19 19 19 20
>
>I understand that neither of these two extremes may be the best when
>it comes to orchestration overhead and load balancing. Instead, the
>best might be somewhere in-between, e.g.
>
>(C) Each worker, called multiple times, processing multiple elements:
>
>> nbrOfElements <- 97
>> nbrOfWorkers <- 5
>> nbrOfChunks <- nbrOfElements / nbrOfWorkers
>> sapply(parallel:::splitList(1:nbrOfElements, nbrOfChunks), length)
> [1] 5 5 5 5 4 5 5 5 5 5 4 5 5 5 5 4 5 5 5 5
>
>However, there are multiple alternatives between the two extremes, e.g.
>
>> nbrOfChunks <- scale * nbrOfElements / nbrOfWorkers
>
>So, is there a reason why you argue for scale = 1.0 to be the optimal?
>
>FYI, In future.apply::future_lapply(X, FUN, ...) there is a
>'future.scheduling' scale factor(*) argument where default
>future.scheduling = 1 corresponds to (B) and future.scheduling = +Inf
>to (A).  Using future.scheduling = 4 achieves the amount of
>load-balancing you propose in (C).   (*) Different definition from the
>above 'scale'. (Disclaimer: I'm the author)
>
>/Henrik
>
>On Mon, Feb 19, 2018 at 10:21 AM, Christian Krause
> wrote:
>> Dear R-Devel List,
>>
>> I have installed R 3.4.3 with the patch applied on our cluster and
>ran a *real-world* job of one of our users to confirm that the patch
>works to my satisfaction. Here are the results.
>>
>> The original was a series of jobs, all essentially doing the same
>stuff using bootstrapped data, so for the original there is more data
>and I show the arithmetic mean with standard deviation. The
>confirmation with the patched R was only a single instance of that
>series of jobs.
>>
>> ## Job Efficiency
>>
>> The job efficiency is defined as (this is what the `qacct-efficiency`
>tool below does):
>>
>> ```
>> efficiency = cputime / cores / wallclocktime * 100%
>> ```
>>
>> In simpler words: how well did the job utilize its CPU cores. It
>shows the percentage of time the job was actually doing stuff, as
>opposed to the difference:
>>
>> ```
>> wasted = 100% - efficiency
>> ```
>>
>> ... which, essentially, tells us how much of the resources were
>wasted, i.e. CPU cores just idling, without being used by anyone. We
>care a lot about that because, for our scientific computing cluster,
>wasted resources is like burning money.
>>
>> ### original
>>
>> This is the entire series from our job accounting database, filteres
>the successful jobs, calculates efficiency and then shows the average
>and standard deviation of the efficiency:
>>
>> ```
>> $ qacct -j 4433299 | qacct-success | qacct-efficiency | meansd
>> n=945 ∅ 61.7276 ± 7.78719
>> ```
>>
>> This is the entire series from our job accounting database, filteres

Re: [Rd] [parallel] fixes load balancing of parLapplyLB

[Henrik Bengtsson]

Hi, I'm trying to understand the rationale for your proposed amount of
splitting and more precisely why that one is THE one.

If I put labels on your example numbers in one of your previous post:

 nbrOfElements <- 97
 nbrOfWorkers <- 5

With these, there are two extremes in how you can split up the
processing in chunks such that all workers are utilized:

(A) Each worker, called multiple times, processes one element each time:

> nbrOfElements <- 97
> nbrOfWorkers <- 5
> nbrOfChunks <- nbrOfElements
> sapply(parallel:::splitList(1:nbrOfElements, nbrOfChunks), length)
 [1] 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
[30] 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
[59] 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
[88] 1 1 1 1 1 1 1 1 1 1


(B) Each worker, called once, processes multiple element:

> nbrOfElements <- 97
> nbrOfWorkers <- 5
> nbrOfChunks <- nbrOfWorkers
> sapply(parallel:::splitList(1:nbrOfElements, nbrOfChunks), length)
[1] 20 19 19 19 20

I understand that neither of these two extremes may be the best when
it comes to orchestration overhead and load balancing. Instead, the
best might be somewhere in-between, e.g.

(C) Each worker, called multiple times, processing multiple elements:

> nbrOfElements <- 97
> nbrOfWorkers <- 5
> nbrOfChunks <- nbrOfElements / nbrOfWorkers
> sapply(parallel:::splitList(1:nbrOfElements, nbrOfChunks), length)
 [1] 5 5 5 5 4 5 5 5 5 5 4 5 5 5 5 4 5 5 5 5

However, there are multiple alternatives between the two extremes, e.g.

> nbrOfChunks <- scale * nbrOfElements / nbrOfWorkers

So, is there a reason why you argue for scale = 1.0 to be the optimal?

FYI, In future.apply::future_lapply(X, FUN, ...) there is a
'future.scheduling' scale factor(*) argument where default
future.scheduling = 1 corresponds to (B) and future.scheduling = +Inf
to (A).  Using future.scheduling = 4 achieves the amount of
load-balancing you propose in (C).   (*) Different definition from the
above 'scale'. (Disclaimer: I'm the author)

/Henrik

On Mon, Feb 19, 2018 at 10:21 AM, Christian Krause
 wrote:
> Dear R-Devel List,
>
> I have installed R 3.4.3 with the patch applied on our cluster and ran a 
> *real-world* job of one of our users to confirm that the patch works to my 
> satisfaction. Here are the results.
>
> The original was a series of jobs, all essentially doing the same stuff using 
> bootstrapped data, so for the original there is more data and I show the 
> arithmetic mean with standard deviation. The confirmation with the patched R 
> was only a single instance of that series of jobs.
>
> ## Job Efficiency
>
> The job efficiency is defined as (this is what the `qacct-efficiency` tool 
> below does):
>
> ```
> efficiency = cputime / cores / wallclocktime * 100%
> ```
>
> In simpler words: how well did the job utilize its CPU cores. It shows the 
> percentage of time the job was actually doing stuff, as opposed to the 
> difference:
>
> ```
> wasted = 100% - efficiency
> ```
>
> ... which, essentially, tells us how much of the resources were wasted, i.e. 
> CPU cores just idling, without being used by anyone. We care a lot about that 
> because, for our scientific computing cluster, wasted resources is like 
> burning money.
>
> ### original
>
> This is the entire series from our job accounting database, filteres the 
> successful jobs, calculates efficiency and then shows the average and 
> standard deviation of the efficiency:
>
> ```
> $ qacct -j 4433299 | qacct-success | qacct-efficiency | meansd
> n=945 ∅ 61.7276 ± 7.78719
> ```
>
> This is the entire series from our job accounting database, filteres the 
> successful jobs, calculates efficiency and does sort of a histogram-like 
> binning before calculation of mean and standard deviation (to get a more 
> detailed impression of the distribution when standard deviation of the 
> previous command is comparatively high):
>
> ```
> $ qacct -j 4433299 | qacct-success | qacct-efficiency | meansd-bin -w 10 | 
> sort -gk1 | column -t
> 10  -  20  ->  n=3∅  19.216667   ±  0.9112811494447459
> 20  -  30  ->  n=6∅  26.418  ±  2.665996374091058
> 30  -  40  ->  n=12   ∅  35.115834   ±  2.8575783082671196
> 40  -  50  ->  n=14   ∅  45.35285714285715   ±  2.98623361591005
> 50  -  60  ->  n=344  ∅  57.114593023255814  ±  2.1922005551774415
> 60  -  70  ->  n=453  ∅  64.29536423841049   ±  2.8334788433963856
> 70  -  80  ->  n=108  ∅  72.95592592592598   ±  2.5219474143639276
> 80  -  90  ->  n=5∅  81.526  ±  1.2802265424525452
> ```
>
> I have attached an example graph from our monitoring system of a single 
> instance in my previous mail. There you can see that the load balancing does 
> not actually work, i.e. same as `parLapply`. This reflects in the job 
> efficiency.
>
> ### patch applied
>
> This is the single instance I used to confirm that the patch works:
>
> ```
> $ qacct -j 4562202 | qacct-efficiency
> 

Re: [Rd] [parallel] fixes load balancing of parLapplyLB

[Christian Krause]

Dear R-Devel List,

I have installed R 3.4.3 with the patch applied on our cluster and ran a 
*real-world* job of one of our users to confirm that the patch works to my 
satisfaction. Here are the results.

The original was a series of jobs, all essentially doing the same stuff using 
bootstrapped data, so for the original there is more data and I show the 
arithmetic mean with standard deviation. The confirmation with the patched R 
was only a single instance of that series of jobs.

## Job Efficiency

The job efficiency is defined as (this is what the `qacct-efficiency` tool 
below does):

```
efficiency = cputime / cores / wallclocktime * 100%
```

In simpler words: how well did the job utilize its CPU cores. It shows the 
percentage of time the job was actually doing stuff, as opposed to the 
difference:

```
wasted = 100% - efficiency
```

... which, essentially, tells us how much of the resources were wasted, i.e. 
CPU cores just idling, without being used by anyone. We care a lot about that 
because, for our scientific computing cluster, wasted resources is like burning 
money.

### original

This is the entire series from our job accounting database, filteres the 
successful jobs, calculates efficiency and then shows the average and standard 
deviation of the efficiency:

```
$ qacct -j 4433299 | qacct-success | qacct-efficiency | meansd
n=945 ∅ 61.7276 ± 7.78719
```

This is the entire series from our job accounting database, filteres the 
successful jobs, calculates efficiency and does sort of a histogram-like 
binning before calculation of mean and standard deviation (to get a more 
detailed impression of the distribution when standard deviation of the previous 
command is comparatively high):

```
$ qacct -j 4433299 | qacct-success | qacct-efficiency | meansd-bin -w 10 | sort 
-gk1 | column -t
10  -  20  ->  n=3∅  19.216667   ±  0.9112811494447459
20  -  30  ->  n=6∅  26.418  ±  2.665996374091058
30  -  40  ->  n=12   ∅  35.115834   ±  2.8575783082671196
40  -  50  ->  n=14   ∅  45.35285714285715   ±  2.98623361591005
50  -  60  ->  n=344  ∅  57.114593023255814  ±  2.1922005551774415
60  -  70  ->  n=453  ∅  64.29536423841049   ±  2.8334788433963856
70  -  80  ->  n=108  ∅  72.95592592592598   ±  2.5219474143639276
80  -  90  ->  n=5∅  81.526  ±  1.2802265424525452
```

I have attached an example graph from our monitoring system of a single 
instance in my previous mail. There you can see that the load balancing does 
not actually work, i.e. same as `parLapply`. This reflects in the job 
efficiency.

### patch applied

This is the single instance I used to confirm that the patch works:

```
$ qacct -j 4562202 | qacct-efficiency
97.36
```

The graph from our monitoring system is attached. As you can see, the load 
balancing works to a satisfying degree and the efficiency is well above 90% 
which was what I had hoped for :-)

## Additional Notes

The list used in this jobs `parLapplyLB` is 5812 elements long. With the 
`splitList`-chunking from the patch, you'll get 208 lists of about 28 elements 
(208 chunks of size 28). The job ran on 28 CPU cores and had a wallclock time 
of 120351.590 seconds, i.e. 33.43 hours. Thus, the function we apply to our 
list takes about 580 seconds per list element, i.e. about 10 minutes. I 
suppose, for that runtime, we would get even better load balancing if we would 
reduce the chunk size even further, maybe even down to 1, thus getting our 
efficiency even closer to 100%.

Of course, for really short-running functions, a higher chunk size may be more 
efficient because of the overhead. In our case, the overhead is negligible and 
that is why the low chunk size works really well. In contrast, for smallish 
lists with short-running functions, you might not even need load balancing and 
`parLapply` suffices. It only becomes an issue, when the runtime of the 
function is high and / or varying.

In our case, the entire runtime of the entire series of jobs was:

```
$ qacct -j 4433299 | awk '$1 == "wallclock" { sum += $2 } END { print sum, 
"seconds" }'
4.72439e+09 seconds
```

Thats about 150 years on a single core or 7.5 years on a 20 core server! Our 
user was constantly using about 500 cores, so this took about 110 days. If you 
compare this to my 97% efficiency example, the jobs could have been finished in 
75 days instead ;-)

## Upcoming Patch

If this patch gets applied to the R code base (and I hope it will :-)) my 
colleague and I will submit another patch that adds the chunk size as an 
optional parameter to all off the load balancing functions. With that 
parameter, users of these functions *can* decide for themselves which chunk 
size they prefer for their code. As mentioned before, the most efficient chunk 
size depends on the used functions runtime, which is the only thing R does not 
know and users really should be allowed to specify explicitly. The default of 
this new optional parameter would be 

[Rd] [parallel] fixes load balancing of parLapplyLB

[Christian Krause]

Dear R-Devel List,

**TL;DR:** The function **parLapplyLB** of the parallel package has 
[reportedly][1] (see also attached RRD output) not
been doing its job, i.e. not actually balancing the load. My colleague Dirk 
Sarpe and I found the cause of the problem
and we also have a patch to fix it (attached). A similar fix has also been 
provided [here][2].

[1]: 
https://stackoverflow.com/questions/38230831/why-does-parlapplylb-not-actually-balance-load
[2]: https://bugs.r-project.org/bugzilla3/show_bug.cgi?id=16792


## The Call Chain

First, we traced the relevant R function calls through the code, beginning with 
`parLapplyLB`:

1.  **parLapplyLB:** clusterApply.R:177, calls **splitList**, then 
**clusterApplyLB**
2.  **splitList:** clusterApply.R:157
3.  **clusterApplyLB:** clusterApply.R:87, calls **dynamicClusterApply**
4.  **dynamicClusterApply:** clusterApply.R:39


## splitList

We used both our whiteboard and an R session to manually *run* a few examples. 
We were using lists of 100 elements and 5
workers. First, lets take a look at **splitList**:

```r
> sapply(parallel:::splitList(1:100, 5), length)
[1] 20 20 20 20 20

> sapply(parallel:::splitList(1:97, 5), length)
[1] 20 19 19 19 20

> sapply(parallel:::splitList(1:97, 20), length)
 [1] 5 5 5 5 4 5 5 5 5 5 4 5 5 5 5 4 5 5 5 5
```

As we can see in the examples, the work is distributed as equally as possible.


## dynamicClusterApply

**dynamicClusterApply** works this way (simplified):

1.  it first gives a chunk to each worker
2.  once a worker comes back with the result, it is given the next chunk

**This is the important part:** As long as there are **more** chunks than 
workers, there will be load balancing. If
there are fewer chunks than workers, each worker will get **at most one chunk** 
and there is **no** load balancing.


## parLapplyLB

This is how **parLapplyLB** splits the input list (with a bit of refactoring, 
for readability):

```r
parLapplyLB <- function(cl = NULL, X, fun, ...)
{
cl <- defaultCluster(cl)

chunks <- splitList(X, length(cl))

do.call(c,
clusterApplyLB(cl, x = chunks, fun = lapply, fun, ...),
quote = TRUE)
}
```

For our examples, the chunks have these sizes:

```r
> sapply(parallel:::splitList(1:100, 5), length)
[1] 20 20 20 20 20
```

There we have it: 5 chunks. 5 workers. With this work distribution, there can't 
possibly be any load balancing, because
each worker is given a single chunk and then it stops working because there are 
no more chunks.

Instead, **parLapplyLB** should look like this (patch is attached):

```r
parLapplyLB <- function(cl = NULL, X, fun, ...)
{
cl <- defaultCluster(cl)

chunkSize <- max(length(cl), ceiling(length(X) / length(cl)))

chunks <- splitList(X, chunkSize)

do.call(c,
clusterApplyLB(cl, x = chunks, fun = lapply, fun, ...),
quote = TRUE)
}
```

Examples with a cluster of 5 workers:

```r
# length(cl) < length(X)
> sapply(parallel:::splitList(1:100, ceiling(100 / 5)), length)
 [1] 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5

# length(cl) >= length(X)
> sapply(parallel:::splitList(1:4, 4), length)
[1] 1 1 1 1
# one worker idles here, but we can't do better than that
```

With this patch, the number of chunks is larger than the number of workers, if 
possible at all, and then load balancing
should work.

Best Regards

-- 

Christian Krause

Scientific Computing Administration and Support



Phone: +49 341 97 33144

Email: christian.kra...@idiv.de



German Centre for Integrative Biodiversity Research (iDiv) Halle-Jena-Leipzig

Deutscher Platz 5e

04103 Leipzig

Germany



iDiv is a research centre of the DFG – Deutsche Forschungsgemeinschaft

iDiv ist eine zentrale Einrichtung der Universität Leipzig im Sinne des § 92 
Abs. 1 SächsHSFG und wird zusammen mit der
Martin-Luther-Universität Halle-Wittenberg und der 
Friedrich-Schiller-Universität Jena betrieben sowie in Kooperation
mit dem Helmholtz-Zentrum für Umweltforschung GmbH – UFZ. Beteiligte 
Kooperationspartner sind die folgenden
außeruniversitären Forschungseinrichtungen: das Helmholtz-Zentrum für 
Umweltforschung GmbH - UFZ, das
Max-Planck-Institut für Biogeochemie (MPI BGC), das Max-Planck-Institut für 
chemische Ökologie (MPI CE), das
Max-Planck-Institut für evolutionäre Anthropologie (MPI EVA), das 
Leibniz-Institut Deutsche Sammlung von Mikroorganismen
und Zellkulturen (DSMZ), das Leibniz-Institut für Pflanzenbiochemie (IPB), das 
Leibniz-Institut für Pflanzengenetik und
Kulturpflanzenforschung (IPK) und das Leibniz-Institut Senckenberg Museum für 
Naturkunde Görlitz (SMNG). USt-IdNr. DE