[LLVMdev] Automating Diagnostic Instrumentation

[Vikram S. Adve]

Reid,

Adding this kind of instrumentation pass would be very valuable in 
LLVM.  We don't have any such thing at present, and there could be 
multiple uses for it.

Joel Stanley did an MS thesis last year that could complement this kind 
of pass nicely.  Joel's thesis was on dynamic performance 
instrumentation guided by explicit queries within the application (I 
will forward you a copy).  Think about it as two things:

(1) A simple performance query language that allows queries to be 
embedded within an application (e.g., "how many L2 cache misses does 
the loop nest labeled X incur, in those instances when a moving average 
cost for function Y is more than 3x of long-term average, i.e., the 
function Y has been unusually slow").  The language allows the user to 
define arbitrary "metrics," to specify routines that can be used to 
measure those metrics, and then to query those metrics for arbitrary 
points or intervals within the application.  A number of common 
computational performance metrics and their measurement routines are 
predefined, e.g., elapsed user/total/system time, L1/L2 cache misses, 
TLB misses.  Many more predefined ones can be added, including OS, 
networking, and other kinds of metrics.

The query language is actually implemented as a simple API.  Joel wrote 
an LLVM pass that recognizes calls to this API, and replaces them with 
initial calls to his runtime system.

(2) A sophisticated runtime system that dynamically inserts and removes 
calls to instrumentation routines that actually do the measurement.  
This is driven by the requirements of the actual queries, e.g., for the 
example query above, you would insert instrumentation for L2 cache 
misses around loop nest X.

Your automatic pass could potentially use Joel's runtime support to do 
the actual work of inserting and removing instrumentation -- the pass 
would only have to insert the appropriate queries in our query 
"language" API.

Caveat: the runtime instrumentation library has only been lightly 
tested and isn't robust yet.

--Vikram
http://www.cs.uiuc.edu/~vadve
http://llvm.cs.uiuc.edu/

On Apr 7, 2004, at 11:32 AM, Reid Spencer wrote:

> Dear List,
>
> I have some questions about some passes for LLVM that I'm thinking 
> about
> but first I need to give a little background ...
>
> One of the things I do in my "day job" (large IT system performance
> analysis) is to figure out where to place diagnostic instrumentation
> within an application. The instrumentation I'm talking about here is 
> the
> kind that is inserted into application code to capture things like
> wall-clock latency, cpu time, number of i/o requests, etc.  For 
> accurate
> diagnosis, it is often necessary to collect that kind of data about
> every occurrence of some event or event pair (like entry/exit of a
> function).  However, this kind of instrumentation can have very high
> overhead if used liberally. When an application is under heavy load, it
> is critical to select the instrumentation points correctly so as to
> minimize overhead and maximize the utility of the information provided.
> Fortunately, there is a way to do this automatically. By constructing a
> static call graph of the application, it is possible to discover the
> "fan-out" points. Fan-out points are calls in the call chain that are
> called from very few places (typically one) but directly or indirectly
> call many functions. That is, some functions in an application are at
> the top of the call chain (e.g. like "main" which implies all the
> processing of the program) while others are at the bottom (like 
> "strlen"
> which implies no further calls and is a leaf node in the call chain). 
> In
> between these two extremes (root and leaf),  the call chain will fan-in
> (like strlen) and fan-out (like main).  Architecturally, we can speak 
> of
> the fan-out points being the places in the application code where a
> module boundary is being crossed.  By instrumenting these fan-out 
> points
> we can be assured that we're instrumenting things that (a) have
> architectural significance (i.e. good quality of information) and (b)
> imply enough processing that the overhead of instrumentation is
> negligible.
>
> With the above kind of instrumentation in mind, I consider such
> auto-instrumentation as "just another pass" in LLVM. That is, I believe
> LLVM makes it pretty easy to do the call graph analysis, find the
> "fan-out" points, and insert the instrumentation code.  So, given that
> its relatively easy to do this, I have the following questions for the
> list:
>
> (1) Would others find this kind of diagnostic instrumentation useful?
>
> (2) Would it be useful to turn the call graph data ino a pretty picture
> via graphviz (both statically and dynamically) ?
>
> (3) How much of this auto-instrumentation pass is already written in
> existing passes?
>
> (4) Are there other ways to achieve the same ends?
>
> (5) Can someone give me some tips on how to implement this?  It would 
> be
> my first LLVM pass.
>
> On a side note, there are source language constructs for which I'm
> wanting to aggregate the data captured by the instrumentation..  Call
> chains of functions are great but sometimes you can't see the forest 
> for
> the trees because the information is too dense. What is useful is to
> aggregate the performance data at higher levels of abstraction. For
> example, in a multi-threaded, object-oriented program you might want to
> see aggregation at the process, thread, class, object instance, and
> method levels. In order to instrument for these kinds of aggregated
> views, I need to be able to provide some kind of "context" (process id,
> thread id, class, etc.) in which a data point is captured.  This 
> implies
> that I need the instrumentation pass to understand some things about 
> the
> source-level language. and possibly capture information about the
> environment the instrumentation application will run in.  
> Unfortunately,
> that means that the pass would either become language specific or
> environment specific because I would have to ensure that the source
> language compiler added the necessary constructs to the generated code
> to provide the contextual information needed by the pass.  This raises 
> a
> couple questions more questions:
>
> (1) Is there be a general mechanism to communicate information between
> source language compiler and LLVM pass? If there isn't, should there
> be?  In the case I describe above it would be *highly* useful (IMO) to
> have the source language compiler provide source level information for 
> a
> language independent pass to use later.
>
> (2) Is there any existing mechanism in LLVM for providing (1) directly?
> What I'm thinking of is some kind of API  that the source language
> compiler can use to add addtional information that any subsequent pass
> might need to use.
>
> Reid.