timecraft: add profile resource #33
Conversation
Signed-off-by: Achille Roussel <[email protected]>
Signed-off-by: Achille Roussel <[email protected]>
|
I added a |
Signed-off-by: Achille Roussel <[email protected]>
Signed-off-by: Achille Roussel <[email protected]>
| // Details about a function call. | ||
| table FunctionCall { | ||
| // State of the WebAssembly stack before and after the function was called. | ||
| // The first {param_count} values are the input parameters, and the remaining | ||
| // values are the return values. | ||
| stack:[ulong]; | ||
| // Captured sections of the WebAssembly module's linear memory, stored | ||
| // contiguously and indexed by {memory_access}. | ||
| memory:[ubyte]; | ||
| // Ordered collection of memory reads and writes made by the function. | ||
| memory_access:[MemoryAccess]; | ||
| } | ||
|
|
||
| // MemoryAccess represents the capture of a section of memory. | ||
| struct MemoryAccess { | ||
| // Byte offset in the WebAssembly module's linear memory where the memory | ||
| // access starts. | ||
| offset:uint; | ||
| // Byte offset into {FunctionCall.memory}. The length of the captured memory | ||
| // can be derived by comparing the {index_offset} with that of the next | ||
| // {FunctionCall.memory_access}. The length of the final memory region can | ||
| // be derived by comparing the {index_offset} with the length of | ||
| // {FunctionCall.memory}. | ||
| index_offset:uint; | ||
| } | ||
|
|
There was a problem hiding this comment.
This is unrelated to the actual change but was unused.
Signed-off-by: Achille Roussel <[email protected]>
Signed-off-by: Achille Roussel <[email protected]>
| Type: desc.Annotations["timecraft.profile.type"], | ||
| StartTime: human.Time(startTime), | ||
| Duration: human.Duration(endTime.Sub(startTime)), | ||
| Size: human.Bytes(desc.Size), |
There was a problem hiding this comment.
Not sure if this is possible, but are there high-level summary metrics we can derive from the profiles, either immediately accessible or something we can aggregate once and cache? For example, it'd be cool to see avg/max CPU utilization, or avg/max RSS, when listing profiles so that you can see basic trends over time, and see anomalies sticking out so you know what to drill down into.
There was a problem hiding this comment.
I like this idea! I'm going to think about it some more.
CPU utilization is interesting, we can compute I/O wait by summing the time spent in stack traces that end in fd_read/fd_write, and ratio that to the total time to estimate CPU usage.
There are a few things that still need adjustments in the CPU profile, the timing is not very precise because we only increment the monotonic clock when we see a record of host function. We can't use the real-time directly because the CPU on which the profile is generated may be different than the one the program was running, but I think we can create a model which works by being a little clever.
I'll look at those enhancements in follow-ups, thanks for the suggestion!
timecraft: add export command
This PR adds a new resource type for performance profiles; they get recorded in the object store and get discovered like other resources.
Still fleshing out how to get pprof launched off of the recorded profiles (so still WIP).