Immix on GHC Summer of Code report #11

My project.

This post assumes that the reader has read my last post.

I’ve been trying to find a reasonable method (which computer to use, single user or multi user) to benchmark for some time, and now an e-mail from Simon Marlow calmed me down. He suggests that in this stage it’s not necessary to run a long benchmark, but only a short one, that will be enought to have an idea about what each change is causing to performance.

For some time I was struggling with the benchmarks to get a reasonable result. I’ve started more than once to write a blog post analysing the results I’ve got, until in the middle of the post I notice the results are not consistent. This usually happens when there’s a big distortion in benchmarks that shouldn’t present no change. For instance, in a patch that only changes code that is executed with +RTS -w -RTS, there was a difference of 50% on execution time in default mode, that is, without +RTS -w -RTS.

I think that only disabling CPU scaling and running a small number of tests, which make it easier to reproduce a possibly inconsistent result, will be enough to get a good data.

I’ve made a stack of patches, splitting each part of my work. The stack does not reflect my preferences between the patches, that is, the complete stack applied is not what I think that is the best change to GHC. It’s only a set of patches applied linearly, in a way that it’s easy to see the impact in performance of each change. The patches are:

Don’t check for swept blocks in -DS.

The checkHeap function assumed the allocated part of the block contained only alive objects and slops. This was not true for blocks that are collected using mark sweep. The code in this patch skip the test for this kind of blocks.

This patch is only needed to make it possible to run the binary with the parameters +RTS -w -DS -RTS. I’ve already explained this patch. I didn’t measure the difference in performance before this patch, since it only changes code that is not going to be executed in the benchmarks, since none of them are runned with -DS.

Immix: allocate and free memory on lines.

This is the main patch with the initial implementation of Immix. As it is the initial implementation, it has it’s problems, that will be treated in the following patches. The comparison was done using 3 programs from nofib/gc: fibheaps, fulsom and constraints, as suggested by Simon Marlow. Two tests were made for each version of the compiler code: one using the default GC strategie (copy-collection with mark-compact if there’s few heap space left), another one using mark-sweep (or Immix after my changes). When comparing the default GC strategie between before and after this change, the program got 0.4% slower, with the collection of the generation 0 being 0.1% slower and the collection of the generation 1 0.4% slower. Comparing the mark-sweep with Immix, it got 13.7% slower, 15% in GC0 and 7.2% in GC1. The change in memory used is irrelevant.

Chose between allocating in line or blocks in todo_block_full

The first improvement I’ve made is to change the place where the allocation was searching for free lines. Most of the changes were done in the function alloc_for_copy(), in rts/sm/Evac.c. This is bad because this function is called a lot of times, and should be kept fast. So I moved them to ctodo_block_full() in rts/sm/GCUtils.c. It got about 9.9% and 9.6% faster in default and sweep, which makes it 1% faster and 2.8% slower than in the original code.

Improvements in sweep()

I was bothered that the code I’ve written to free memory in lines was messy, and I thought it could be simplified and maybe even turn faster. The general execution time increased 0.4%, but the GC1 time, which is the only one that should be affected, was reduced by 0.6%. In the default GC strategie, there was also a 0.2% increase in time, so I guess this is inside the margin of change. I have to investigate better on this one, but it seems to have improved slightly the performance.

Line before inscresing block size.

In a conversatio with Simon Marlow, we thought about these two options: one is to search for a line after trying to increase the block size, and that’s what I did in Chose between allocating in line or blocks in todo_block_full; the other is trying to search a line first, and just increase the block size if a block (and not a line) was being used for allocation and there is no line available. With this patch I’ve changed it to use the second one. Using sweep, the GC time got 0.5% slower, and with default it got 0.1%.

rts/sm/Sweep.c: Mark all BF_MEDIUM blocks as BF_FRAGMENTED

In my first implementation, the blocks that contain objects bigger than a line (of medium size) are marked with BF_MEDIUM, and are treated as the usual block in mark-sweep: if the block is empty, it’s freed; if it’s very fragmented, it’s marked to be collected using copy-collection. In this patch, instead of making it use the rules of mark-sweep, I just mark it as fragmented and make it be collected by copy-collection. This made the GC0 time be reduced by 3%, but caused an increase of 509.7% in the GC1 time. This was also the first time the memory used was reduced, by 25.4%. The results in the default mode are insignificant. What it seems to me is that the copy-collection code is more efficient with memory usage, but also more slow. Marking objects as BF_FRAGMENTED make this algorithm be used even for full objects.

rts/sm/Sweep.c: Don’t use 3/4 heuristics to mark as BF_FRAGMENTED

The mark-sweep algorithm considered that an object was fragmented if more than 1/4 of the word groups are completely unmarked. I kept using this heuristic, and this change removes it, making more use of the Immix, and less use of copy-collection. This didn’t change the memory used, and made the code 10.4% slower in sweep. In default GC mode, there was no relevant difference, as expected.

Allocate in lines in minor GCs.

In the initial implementation I was only allocating on lines only in major GCs, because I needed the mark stack, which was only available in these kind of allocations. In this change, I created the mark stack on all GCs, and used the allocation on lines. The results for the default mode are insignificant, and the code got 9.0% faster in sweep, using the same memory.


I’ve made a selection with the presumably better patches, which are: Don’t check for swept blocks in -DS, Immix: allocate and free memory on lines, Chose between allocating in line or blocks in todo_block_full, Improvements in sweep() and Allocate in lines in minor GCs. This is an attempt to achieve the best set of patches possible, to see how it improved the original code. Comparing the default strategie for this selection and the original code, it got 3.7% slower. Comparing both sweep, it got 3.9% slower. Comparing the original default with the final sweep, it got 4.2% slower and uses 4.7% more memory. There’s a lot of room for improvement, and I’m willing to hear suggestions of what I could change in the code to achieve this improvement.

The results are available. Every time I had doubt about a comparison, I’d run both versions again to check. This is way there are some backup files. The data presented here is not from all the most recent versions of each measurement, but from the ones that I thought were more similar in conditions.


2 Respostas to “Immix on GHC Summer of Code report #11”

  1. Immix on GHC Summer of Code report #12 / DebConf / Debian Day BH « Blog do Marcot Says:

    […] Blog do Marcot « Immix on GHC Summer of Code report #11 […]

  2. Immix on GHC Summer of Code final report « Blog do Marcot Says:

    […] This part of this post assumes that the reader has read my last post. […]

Deixe um comentário

Preencha os seus dados abaixo ou clique em um ícone para log in:

Logotipo do

Você está comentando utilizando sua conta Sair /  Alterar )

Foto do Google+

Você está comentando utilizando sua conta Google+. Sair /  Alterar )

Imagem do Twitter

Você está comentando utilizando sua conta Twitter. Sair /  Alterar )

Foto do Facebook

Você está comentando utilizando sua conta Facebook. Sair /  Alterar )

Conectando a %s

%d blogueiros gostam disto: