Objet : Developers list for StarPU
Archives de la liste
- From: Konstantin Sozykin <Ko.Sozykin@skoltech.ru>
- To: "starpu-devel@inria.fr" <starpu-devel@inria.fr>
- Cc: Aleksandr Mikhalev <Al.Mikhalev@skoltech.ru>, Roman Schutski <R.Schutski@skoltech.ru>, Georgii Ovchinnikov <G.Ovchinnikov@skoltech.ru>
- Subject: [starpu-devel] Customization/Optimization for Scheduling Methods
- Date: Fri, 14 Jul 2023 14:52:26 +0000
- Accept-language: ru-RU, en-US
- Authentication-results: mail3-smtp-sop.national.inria.fr; spf=None smtp.pra=Ko.Sozykin@skoltech.ru; spf=Pass smtp.mailfrom=Ko.Sozykin@skoltech.ru; spf=None smtp.helo=postmaster@srv-ksmg-01.skoltech.ru
- Ironport-data: A9a23:pmileqwoFNLaheSin596t+f2wirEfRIJ4+MujC+fZmUNrF6WrkUPm jEcCGiBb6yMYDShet0kPIznpkxTucXXn9drHAo9qlhgHilAwSbnLYTAfx2oZ0t+DeWaERk5t 51GAjXkBJppJpMJjk71atANlVEliefSAOCU5NfsYkhZXRVjRDoqlSVtkus4hp8AqdWiCmthg /uryyHkEAHjg2cc3l48sfrZ80sw5aiq4Vv0g3RnDRx1lA+G/5UqJMlHTU2BByOQapVZGOe8W 9HCwNmRlo8O10pF5nuNy94XQ2VSKlLgFVDmZkl+B8BOtiN/Shkaic7XAhazhXB/0F1ll/gpo DlEWAfZpQ0BZsUgk8xFO/VU/r0X0QSrN9YrLFDm2fF/wXEqfFOwkrZBL2YXN7YzxcYmEVEX2 t09OW0SO0Xra+KemNpXS8F3nNpmL9LuM8ZH4zd7zjbdEfstQdbZXs0m5/cBg2d23JseW6+EO 4xAMmYHgBfoO3WjPn80D9QFm/W5hXj5WyZbr1aJqKE3pXLOpOB0+Oa0bIOKI4DXLSlTtmG2h m/ooUPBOD0DaYy02DCc+CiiodaayEsXX6pJTuHgr6Y06LGJ/UQIFBQMTUb+rfSnh0qWQMNaM 0VS+yw0rKF0+lbDczXmdwaivjuHpRsTHoIJVfY57wWRx6HUpRyEboQZctJfQIB39+AVYh0g7 XS2v4L5NWVAkZiEFG3Io994sgiOESQSKGYDYwoNQg0E/8TvrekPYvTnE44L/Emd04SdJN3g/ 9yZhHVj2OlD16bnw43gog6e3GrESo3hF1Ztvm3qsnSZAhRRXrPNi2aAzF/A5vFfRGpyZgDb4 yRfx6ByAMgrCoqGkzPlfQngNKqz+bOGKzjTxAY2WYIi8DCx9nWjO5pNiN2fGKuLGptbEdMKS BWL0e+02HO1FCHxBUOQS9npY/nGNYC6SbzYugn8N7KimKRZeg6d5z1JbkWNxW3rm0VEufhha cbBLp30VixKVf0PIN+KqwE1jOdDKscWmDq7eHwH503PPUe2PyLNFedt3KWmNb9ohE97nOkl2 40Fa5PTk0s3vBzWYDLcuYsaNkwHN2Q2AfjLRz9/K4a+zv5dMDh5UZf5mOpxE6Q7xP49vrmTo RmAtrpwlAaXaYvvcl/ROxiOqdrHAf5CkJ7MFXVyZgn3gyZ8PtrHAWV2X8JfQITLPddLlZZcJ 8Tpse3ZahiWYmSXoWYuftPmoZZ8dR+mowuLMmD3KHI8ZpNsDUiBsNPtYgKlpmFEAzuVpPkOh eSq9jraZp4fGCVkLsLdM8y0w32L4HMypeNVXmnzGOd1RnnCyoZRBhbKvqcFGP1UcRTn7RmG5 jmSGiYd9LXsoZdq0dznhpKkjoaOEslsFxALBGX007S/HnTL8lGC2q5Fav6DJhrGZVP3+YKjR ORb9O79O/s5h2R3s5JwPrJo7KAm7f7tmuNq9RtlF3D1cFibMLNsDX2Y181ptKcW5LtmlSape 0CIoP92BK6oPZ77LVsvOwYVVOSP+vUKkD307/5uAkHb5jdyzYWXQ3dpIBiApyxMHoRbaLp/7 78ah/cXzAijhj4BENWM1HlU/lvRCE0wafwss5VCDbL7jgYu9Ep5XqXdLS3IsbWvcNRHN3c4L gCE3JTig6tu/WucUn4RO0WU489jq8Uvgi1a9H4DOFWDpfTdjNAVwhB60Go6XyZV/Dp9wsNxP WlgbG8vAZqS7T1Xj9Rmc02xEFtdGhe5wEjAkVQUn2vkTm2DdG/EHEsiM8mjoWEb9GN9eGBA3 be6kWzKbxfjTPvT7AATB3F3juPFTMNg0DHClOSMPdW3L7NjbRXL2qaRNHc18T35CsYPtWj7j OhN/tcoT5bkNCQV8pYJO6PD2Zs+EBm7dXF/G9d/96Y0HEbZSjG4+R6KD2uTIsptBfj7wXWUO vxUBPBkdkqBjX6VjzUhG6QzDad+n6cp6Po8a7rbHzM6nIXFnAV5kqD713bYtDcwTsRMgPQND NrbVwi/H1y6gVpWnG7wr/d4BFeoXOlcWirC2LGazeZYMbMCr+BmTm8q2JSWoXi+EVVq7jCUj iz5do7UyO1pk5hylavWFpQZCBi/L+7ycM+q8wmDlctEQv2SEMXJtiIT8kLGOSYPN5Qvet1Hr 5a/m/+p41Hk5ZEYC3v4nbuFHIl3vfSCZvJdaJ/LHSMLjBm8V9/JyDpd3WKBcLhiss5Xv+ujT CuGMPqAT8YfAYph9ScEehplMkgvDorsZf3dvgK7lfOHDyYd3SHhLN+K8XzIb3lRRhQXOq/RW xPFhPKz2u929Ih8JgcIJ/VDMa9KJFXOXagHddqouwfBXyPsyhmHt6D5nBUt1SDTBzPWWIzm6 JbCXV7leA70pKjMy8pDvpdvugEMSkxwmvQ0YllX7useZ+pW14LaBb913VQ65pBofujaz4vmI TLWaWRnUXy7RjdDdQnx69SlRBr36ini/DvmDmRBwq9WQ37e6EC87H9J7SRn52t7cTul1vzPx RQ25CjrJhboqn12bb972xF46NuLAtvBxXUM5U32k4roHn7yxFnMOGNJRGJwaMAMLy0BeIgn6 4T4qaCojXxXkXLMLPs=
- Ironport-hdrordr: A9a23:SZfk4K+UFA6e2yGH92Buk+F7d71zdoMgy1knxilNoERuA6qlf8 DHppQmPGzP+U8ssRAb6KC90cy7K080mqQFhbX5UY3NYOCEghraEGgB1/qE/9SIIUSXnYMwup uIGJIOa+EYeGIK6PoSlzPIcerIruPnzImYwcXl4zNEShxrUbhp60NYDQmce3cGDXgPdNFJXu vQl4Y36EvvCDVnLrX5OlA1G8r4i5nik4/nWwcPA1oC5BSOiz7A0s+MLzGomjUlf3dmxqojy3 HMl0je+ripte29x1v90HLe9J5bgcbm0Z9iA8OL4/JlWAnEu0KLXqgkc7qevAotqOXq0142nL D30mEdFvU2yUn8OkmzvBvWwgHllAwy42L/0zaj8A/eiP28ZggXT+JHmIJiaxPcr3Utp8t91q Uj5RPti7NnSTH7sGDU4MXJThBrjEu1pnRnuuYIg2Y3a/pYVJZh6a4k0Ad8F4oBBSP78oEmF+ crLMbS4p9tABGnRkGcnFNChPSqQ3QrHxedQ0gEtIiw3yVWmn0R9Tpd+OUv2lU72NYUTIRF/u XFKKVpk78LdMUfaqV8A+k9UdK6Y1a9BC7kASaxG3DIUIkBJnL1qp7s4bUz6oiRCc415ap3uq 7kfhdxsnMzQVvnDYmnwIBLmyq9HlmVbHDO8OVxo7B/p73HXbLudQGFV1wqldKt59oFBMmzYY foBLtmR93YCUWrMYFV3zfmU50XCX4QVqQuy4cGcmPLgPjuBMnDsfHaSu/aIP7MFispQWnyG2 YER3zSI8pHh3rbCUPQsVzyYVfCPmT54JJLCaDfu80OwMw1LYtCsg0Ykz2Cl5m2AAwHkJYTUg 9XJ67mqLqyoS2T8XvI6WgBAGsDMm9lpJfbd10PjQgWM3nsebJGgN2DdQlprSa6Dy46bfnzPk p2oEl2w7u+K9ir2Cg+F8jPCBPHs1Ij4FCxa7lZsqqf5f39cpd9MZo8Qqx2EmzwZkdIsDcvil 1mRkstQVLeKTP0lb/Ntu1AOMjvM/tHpCnuCclIqU3erlmH4fszTmYGRVeVIJqqqDdrZQBxqB lU/7UEiLSN3R6kNGs4h40DQQRxQVXSIohjIkC/eIBZ3pDHETsAC1ti2Abq2y3bMVCaj3n6zl aRVhG8aLXIGFYYsndE0rnr609vMmWbcEQYUAEFjaRtUWbdtjJ61vKKeaqpz3HUalwezvoWPC jZZyBXIgRgwbmMpWGosSfHF28rgp0yI+TWDbAudKzd2naxJJaJkq9DH/hI8I1uOMnytPROUe WRfWauXU/FIvJs2RaY4n04PStxpGM/mf3lwhLk6XW1x345D77JKlxgA6oWOc2Y8gHfNoe1+Y Q8j8g4+fG9KXr8cd6Axqe/VU8/Fvr8mx/Nc90V
- Ironport-phdr: A9a23:B3aqSRee6lHszFIgDbNG46ILlGM+L9TLVj580XLHo4xHfqnrxZn+J kuXvawr0AWXG96FoKIcw8Pt8InYEVQa5piAtH1QOLdtbDQizfssogo7HcSeAlf6JvO5JwYzH cBFSUM3tyrjaRsdF8nxfUDdrWOv5jAOBBr/KRB1JuPoEYLOksi7ze+/94PSbglSijewYLx+I RaqoQ7MqsQYnIxuJ7orxBDUuHVIYeNWxW1pJVKXgRnx49q78YBg/SpNpf8v7tZMXqrmcas2S 7xYFykmPHsu5ML3rxnDTBCA6WUaX24LjxdHGQnF7BX9Xpfsriv3s/d21SeGMcHqS70/RC+v5 Ll3RhD2lCgHNiY58GDJhcx2kKJbuw+qqxhmz4LJfI2ZKP9yc6XAdt0YWGVBRN5cWSxfDI6ic oUPE+QPM+VZr4bhqFQDtgGxCRWuBO711jNEmn770Ksn2OohCwHG2wkgEsoLvnTTqtX1L7sSX vqzzKnUzDXDae5d1yr66InJchEhovCMXbRtesfWzEcvEBnJgEiXqY3kPzKV1eUNs26A4up7V uKvjGonqx1vrTezxscgkJPJhp8Mx13C6C52z5o7K8eiR05nfd6rDoFQtyeCOodoX84sTG9lt igmxrAHupO2czUHxIg7yxLBa/KKcIqF7xPnWeuNJTp2i31rdK+hixqu/katyenxW8m03VpXs ydIlMTHuHMV1xHL98SKSPhw8l2/1TuB1A3f8PxILV47mKbBNpIszLE9moIOvUnBHSL6glv6g aGVe0k+5+Sk9evqbq/7qpKTKYN5jBz1PL40lcylG+s4NxADX2iF9uS4073u5Vb5QK9UjvEvk anZq5XaJcsCqqKlHgNZyJ0j6xW4DzeizNQXg2UILFJfdBKdkojlI1DOIPbmAvejm1mgjTVmy vPcMrH/DJjALWLPnKr/cbty8UJQ1QQ+wN9H65JREL4BIfbzWkHrtNzfCx80Kwm6w+H9BNtg0 YMfWGKOAqyYMKPUqlKI/eEvI+6IZI8JpjnyN/gl6+T0jX8kgVMde7Km0oMNaH+kBvRmP1mZY X30j9scCWsKpBYxTPT2iF2eVj5ef22yX74m6TEhEIKmEJ7PSZ2zj7yfwii7GoZbZntGC1CJC XfnbZ+IW/YKaCKII89uiCYIVba7S9xp6Rb7/hTmwqB/M6/Y9zMVsbri0sJ07qvdj1t6oSdoB t6FziSBQn95mksJQSU31eZxuxou5E2E1P1bhPpDCdVV+/JPGjw9J4/fzuoyX/P1HDnGbMaKS VCOWdahBS08RdJ32cJYMBU1IMmrkh2Wh3niOLQSjbHeQcVsqso0vlD0Lsd5kDPd0bU5ykIhW o1JPHGngah2807SAZTImgOXjfXibrwSiQjK8mrL1m+SpAdASgclcqKDZncNd0jXq/zi5UTGU 7KlCfI9L1gJ0taMf5NDccahllBaXLHmMdXabXi2njKTAlCzwaGdb4vsU34e2yHBBUME1RoOr j6dLQZrICCnrirFCSB2U1LiZ0S56e5ltHayVVM51SmRdVMk1KGw9VtN3LqHSPQXwr8PuGE7t l2YBX6b2NTbQ5qFrgtlJ+BHZM8lpUxAzSTfvhB8OZqpK+ZjgEQfekJ5pRGm0RI/EYhGncUwy RFihANvNaKV1k9AfDKEzNjxPLPQMGz74BGobebfxFjf1N+c/qpH5u4/rh3vuwSgF0xq9HsCs ZEd8Hza3ZLYHQkTV7rqUUEz6xVxp/fAf2h15o/Z02FtLbjhqiXLiLdLTKMuzhetecsaMbvRT VCiVZdBV473crBswgP1PXdmdKhI+aU5Pt2rba6D0a+vZ6N7mS6+yH9A+MZ7216N8Cx1TqjJ2 YwEyreWxFjiNX+0gVG/v8TwgY0BaysVGz/16i+iP4lDd65ze64QCmCqP8C8xpNkmtS+PhwQv E7mHF4A1MKzLFCxaBrG2h9B1kURiWekni+kwjd01Ss05Pn6vmSG06HpcxwJPXROTW9pgAL3I IS6uNsdWVChcwkjkBbNCV/S/6FAv+w/KmDSRRwNZC3qNyR5VbP2sLOeYslJ4ZdusCNNUe36b 0rIArL6phIb1WvkEQ48jHgYcHeRvYfonB18oHmeK3JvoXHZP9ttjRvS/93TQ/dN0yFOHXYoz 2OMVh7lZYXvpI7I357Y14L2H3qsTJhSbTXmwcubuS236HcrSRyzkvavm8H2RA0z0Cv1zd5vB kCq5F73Zojm0bj/MPoyJxY4QgWnu4wjQN04y9Bs1/RykTAAi56Y/GQKizL2ONRfguflaWYVA CUM25jT6RTk30tqKjSIwZj4XzOT2JgEBZHyb2UI1yY6980PBr2T6ekOviwzmV2qvAHQZ9Bmn TAZ1fYl6DgAnqta3WhlhjXYGb0UEURCaGbPkFKz6MKvr6RRTH2sdrys2UF+28q8RuLnwEkUS DPyfZEsGjV158N0PQfX0XH93YrjfcHZcdMZshDH2weFleVeL4g90+YbnScyc3yopmUrkqRo6 H4mlYH/poWMLH9hub60EgINfCOgfNsdo3nslfoMx5fLmdn0RdM7Q2tNBsGwBfOwTGBL6qShb V3eVmRm9jHCQua6f0fX6V86/SueTtbybCrRfiBHi484G1GcPBAN2V9EGm9hzthgTFzsnZewO A94/mxDvw+l7EEUmqQxcUCnDgK97E+pcmtmEcfPakANvkcbvRqTa5HGpuNrQ3MIr9v49FLLc zPHIV0SVwRrEgSFAVvnI7Wjt8LY/a6DC+yyNPLfJ6+T4e1YXPOVyZ/9yZNouSuKN96COHVvF bsw3U8mPzgxGsLSnygDRn4MjyyLcsiaoxGxvCxw5sG5+/D2VA+9/pOBTaZUMM5i/Ry7m+GZK qaGjSJ9Y160z7sqwnnFgPga1V8W0GR1ciW1VK8HrWjLRb7RnaleC1gabTlyPY1G9fB02A4FI sPdht7vs9wwxvcoF1dIU0DgkcC1dIQLJW+6LlbOGEeMMvyPOzTKx8j9Za70R6dXia1YsBi5u DDTFEGGXHzLjz7ySxWmKv1BlgmAJgIbu5q4fVM1UC74T9XidhCwMJltkHx+wLE5gG/LKX9JM TV4dBAozPXY5idZj/NjXm1ZuyM6fa/ewHbftrOEbM9N7agOYGw8je9R7XUkxqEA6ShFQKcwg y7Otpt1pEnglOCTyz1hWR4IqzBRhYvNs18xXMeRvpRGR3vA+woAqGuKDBFf7fhiTOboqbFTy 97nj6H0LCtP+NKS59ZWVK22YIqXdWEsNxbkAmueFAweUTuiLn3SnWRGiu/X9WaQod1i8973m 50KVrJQVBovC7lJbyYtVMxHK5BxUDQ+lLedh8Nd/nuyoi7aQ8BCt4zGXPafaR0KAC2air1Ya hAIh6nlf9x73mLTwE9lbkV3l4SMAVeCBLilQwV6dBFxo15I/z0mFyso2EXhdg6m4DkJCKzs9 iM=
- Ironport-sdr: 64b1612e_T4SZrYrR+reRbSs/U890kVAR4i91MU1WZCLR0bUmBvQ7dHa xvjB7XAyseYsjEzq15OUBPvC52krMsf/Liw3Klw==
Dear StarPU Team,
Thank you all for the amazing framework. We are currently working on custom code using StarPU and we discovered that the default schedulers can't handle our custom operations. Consequently, we are considering developing our own optimization- or even a learning-based scheduler. How reasonable is this, given the current development of StarPU?
Additionally, we would like to pose the following inquiries:
Firstly and most importantly, how can we extract the full DAG? Are the *.rec and *.dot files sufficient for this purpose?
Can we access the full DAG in online mode? By this, we mean that we extract the graph during or before computation, send it to an external module, and then return it to StarPU.
We have located functions for exporting linear programs (*.lp and *.mps files) to determine theoretical bounds for execution time. However, as far as we understand, these are only for measuring performance and not for optimizing scheduling. Is our understanding correct?
Thank you for your guidance.
Best regards,
K.
Thank you all for the amazing framework. We are currently working on custom code using StarPU and we discovered that the default schedulers can't handle our custom operations. Consequently, we are considering developing our own optimization- or even a learning-based scheduler. How reasonable is this, given the current development of StarPU?
Additionally, we would like to pose the following inquiries:
Firstly and most importantly, how can we extract the full DAG? Are the *.rec and *.dot files sufficient for this purpose?
Can we access the full DAG in online mode? By this, we mean that we extract the graph during or before computation, send it to an external module, and then return it to StarPU.
We have located functions for exporting linear programs (*.lp and *.mps files) to determine theoretical bounds for execution time. However, as far as we understand, these are only for measuring performance and not for optimizing scheduling. Is our understanding correct?
Thank you for your guidance.
Best regards,
K.
- [starpu-devel] Customization/Optimization for Scheduling Methods, Konstantin Sozykin, 14/07/2023
- Re: [starpu-devel] Customization/Optimization for Scheduling Methods, Samuel Thibault, 17/07/2023
- Re: [starpu-devel] Customization/Optimization for Scheduling Methods, Konstantin Sozykin, 18/07/2023
- Re: [starpu-devel] Customization/Optimization for Scheduling Methods, Samuel Thibault, 17/07/2023
Archives gérées par MHonArc 2.6.19+.