The validity of that advice depends on a lot of factors.  G1 changed the
game a bit for pause times for sure but you can still see larger pause
times than acceptable for some cases.  In any event I agree that we should
be more careful with how we describe heap usage.


On Oct 14, 2016 7:10 PM, "Russell Bateman" <> wrote:

Yeah, I spent a bit of time this morning before posting looking for a magic
8-10Gb advisory and generally for GC gotchas related to larger heap sizes
in the 64-bit world, but couldn't find any. We're using 12Gb right now for
NiFi and haven't noticed any trouble. We vaguely conceive of increasing
this amount in the future as needed as our servers tend to run large
amounts of memory.

The statement yesterday on this thread warning against using that much is
what sent me into Google-it mode. I think this advice is a red herring.


On 10/14/2016 03:03 PM, Corey Flowers wrote:

We actually use heap sizes from 32 to 64Gb for ours but our volumes and
graphs are both extremely large. Although I believe the smaller heap sizes
were a limitation of the garbage collection in Java 7. We also moved to ssd
drives, which did help through put quite a bit. Our systems were actually
requesting the creation and removal of file handles faster than traditional
disks could keep up with (we believe). In addition, unlike with traditional
drives where we tired to minimize caching, we actually forced more disk
caching when we moved to ssds. Still waiting to see the results of that on
our volumes, although it does seemed to have help. Also remember, depending
on how you code them, individual processors can use system memory outside
of the heap. So you need to take that into consideration when designing the

Sent from my iPhone

On Oct 14, 2016, at 1:36 PM, Joe Witt <> wrote:


You can definitely find a lot of material on the Internet about Java heap
sizes, types of garbage collectors, application usage patterns.  By all
means please do experiment with different sizes appropriate for your case.
We're not saying NiFi itself has any problem with large heaps.


On Fri, Oct 14, 2016 at 12:44 PM, Russell Bateman <russell.bateman@> wrote:

> Ali,
> "not recommended to dedicate more than 8-10 GM to JVM heap space" by whom?
> Do you have links/references establishing this? I couldn't find anyone
> saying this or why.
> Russ
> On 10/13/2016 05:47 PM, Ali Nazemian wrote:
> Hi,
> I have another question regarding the hardware recommendation. As far as I
> found out, Nifi uses on-heap memory currently, and it will not try to load
> the whole object in memory. From the garbage collection perspective, it is
> not recommended to dedicate more than 8-10 GB to JVM heap space. In this
> case, may I say spending money on system memory is useless? Probably 16 GB
> per each system is enough according to this architecture. Unless some
> architecture changes appear in the future to use off-heap memory as well.
> However, I found some articles about best practices, and in terms of memory
> recommendation it does not make sense. Would you please clarify this part
> for me?
> Thank you very much.
> Best regards,
> Ali
> On Thu, Oct 13, 2016 at 11:38 PM, Ali Nazemian <>
> wrote:
>> Thank you very much.
>> I would be more than happy to provide some benchmark results after the
>> implementation.
>> Sincerely yours,
>> Ali
>> On Thu, Oct 13, 2016 at 11:32 PM, Joe Witt <> wrote:
>>> Ali,
>>> I agree with your assumption.  It would be great to test that out and
>>> provide some numbers but intuitively I agree.
>>> I could envision certain scatter/gather data flows that could challenge
>>> that sequential access assumption but honestly with how awesome disk
>>> caching is in Linux these days in think practically speaking this is the
>>> right way to think about it.
>>> Thanks
>>> Joe
>>> On Thu, Oct 13, 2016 at 8:29 AM, Ali Nazemian <>
>>> wrote:
>>>> Dear Joe,
>>>> Thank you very much. That was a really great explanation.
>>>> I investigated the Nifi architecture, and it seems that most of the
>>>> read/write operations for flow file repo and provenance repo are random.
>>>> However, for content repo most of the read/write operations are sequential.
>>>> Let's say cost does not matter. In this case, even choosing SSD for content
>>>> repo can not provide huge performance gain instead of HDD. Am I right?
>>>> Hence, it would be better to spend content repo SSD money on network
>>>> infrastructure.
>>>> Best regards,
>>>> Ali
>>>> On Thu, Oct 13, 2016 at 10:22 PM, Joe Witt <> wrote:
>>>>> Ali,
>>>>> You have a lot of nice resources to work with there.  I'd recommend
>>>>> the series of RAID-1 configuration personally provided you keep in mind
>>>>> this means you can only lose a single disk for any one partition.  As long
>>>>> as they're being monitored and would be quickly replaced this in practice
>>>>> works well.  If there could be lapses in monitoring or time to replace 
>>>>> then
>>>>> it is perhaps safer to go with more redundancy or an alternative RAID 
>>>>> type.
>>>>> I'd say do the OS, app installs w/user and audit db stuff, application
>>>>> logs on one physical RAID volume.  Have a dedicated physical volume for 
>>>>> the
>>>>> flow file repository.  It will not be able to use all the space but it
>>>>> certainly could benefit from having no other contention.  This could be a
>>>>> great thing to have SSDs for actually.  And for the remaining volumes 
>>>>> split
>>>>> them up for content and provenance as you have.  You get to make the
>>>>> overall performance versus retention decision.  Frankly, you have a great
>>>>> system to work with and I suspect you're going to see excellent results
>>>>> anyway.
>>>>> Conservatively speaking expect say 50MB/s of throughput per volume in
>>>>> the content repository so if you end up with 8 of them could achieve
>>>>> upwards of 400MB/s sustained.  You'll also then want to make sure you have
>>>>> a good 10G based network setup as well.  Or, you could dial back on the
>>>>> speed tradeoff and simply increase retention or disk loss tolerance.  Lots
>>>>> of ways to play the game.
>>>>> There are no published SSD vs HDD performance benchmarks that I am
>>>>> aware of though this is a good idea.  Having a hybrid of SSDs and HDDs
>>>>> could offer a really solid performance/retention/cost tradeoff.  For
>>>>> example having SSDs for the OS/logs/provenance/flowfile with HDDs for the
>>>>> content - that would be quite nice.  At that rate to take full advantage 
>>>>> of
>>>>> the system you'd need to have very strong network infrastructure between
>>>>> NiFi and any systems it is interfacing with  and your flows would need to
>>>>> be well tuned for GC/memory efficiency.
>>>>> Thanks
>>>>> Joe
>>>>> On Thu, Oct 13, 2016 at 2:50 AM, Ali Nazemian <>
>>>>> wrote:
>>>>>> Dear Nifi Users/ developers,
>>>>>> Hi,
>>>>>> I was wondering is there any benchmark about the question that is it
>>>>>> better to dedicate disk control to Nifi or using RAID for this purpose? 
>>>>>> For
>>>>>> example, which of these scenarios is recommended from the performance 
>>>>>> point
>>>>>> of view?
>>>>>> Scenario 1:
>>>>>> 24 disk in total
>>>>>> 2 disk- raid 1 for OS and fileflow repo
>>>>>> 2 disk- raid 1 for provenance repo1
>>>>>> 2 disk- raid 1 for provenance repo2
>>>>>> 2 disk- raid 1 for content repo1
>>>>>> 2 disk- raid 1 for content repo2
>>>>>> 2 disk- raid 1 for content repo3
>>>>>> 2 disk- raid 1 for content repo4
>>>>>> 2 disk- raid 1 for content repo5
>>>>>> 2 disk- raid 1 for content repo6
>>>>>> 2 disk- raid 1 for content repo7
>>>>>> 2 disk- raid 1 for content repo8
>>>>>> 2 disk- raid 1 for content repo9
>>>>>> Scenario 2:
>>>>>> 24 disk in total
>>>>>> 2 disk- raid 1 for OS and fileflow repo
>>>>>> 4 disk- raid 10 for provenance repo1
>>>>>> 18 disk- raid 10 for content repo1
>>>>>> Moreover, is there any benchmark for SSD vs HDD performance for Nifi?
>>>>>> Thank you very much.
>>>>>> Best regards,
>>>>>> Ali
>>>> --
>>>> A.Nazemian
>> --
>> A.Nazemian
> --
> A.Nazemian

Reply via email to