Speed up transfer of outputs to notebook webviews

[mjbvz]

For microsoft/vscode-jupyter#11031

The VS Buffer for the output items has a short byte length but a very large backing buffer (Uint8Array which is actually a Buffer)

When sending outputs to the webview, we ended up transfering the huge backing buffer instead of the much smaller actual data

We can fix this by creating a Uint8Array copy of the data and then transfering it to the webview

[mjbvz]

@jrieken @alexdima In this instance, first.data.buffer is a Buffer which looks like a Uint8Array. This Buffer is much, much larger than the VSBuffer for first.data

Is there a way to avoid this size mismatch? I suspect it's an optimization to reduce the number of memory allocations

[rebornix]

When we send notebook model dto from EH to Renderer, we merged all buffers into one

vscode/src/vs/workbench/services/extensions/common/rpcProtocol.ts

Lines 857 to 875 in b9202b6

	private static _serializeReplyOKJSONWithBuffers(req: number, res: string, buffers: readonly VSBuffer[]): VSBuffer {
	const resBuff = VSBuffer.fromString(res);
	let len = 0;
	len += MessageBuffer.sizeUInt32; // buffer count
	len += MessageBuffer.sizeLongString(resBuff);
	for (const buffer of buffers) {
	len += MessageBuffer.sizeVSBuffer(buffer);
	}
	const result = MessageBuffer.alloc(MessageType.ReplyOKJSONWithBuffers, req, len);
	result.writeUInt32(buffers.length);
	result.writeLongString(resBuff);
	for (const buffer of buffers) {
	result.writeBuffer(buffer);
	}
	return result.buffer;
	}

I think that's how we end up with multiple buffers sharing the same ArrayBuffer in the Renderer. And then when we send them to Webview, we create a copy of the ArrayBuffer store for every buffer.

[mjbvz]

Thanks @rebornix. This may be the best fix then. Longer term, we could explore using a SharedArrayBuffer to transfer the data without the extra copies

[rebornix]

I'll merge this first to ensure users don't have OOM with images. We should look into this and see if we can avoid the duplication of buffers.